Merge lp:~smspillaz/compiz-core/compiz-core.work_923683 into lp:compiz-core/0.9.5

The diff is big and scary, but its not too complicated. I'll explain each separate part

1 === modified file 'include/core/window.h'
2 --- include/core/window.h 2012-01-20 15:20:44 +0000
3 +++ include/core/window.h 2012-01-31 17:12:59 +0000
4 @@ -398,9 +398,9 @@
5
6 bool hasUnmapReference ();
7
8 - bool resize (XWindowAttributes);
9 + bool resize (const XWindowAttributes &);
10
11 - bool resize (Geometry);
12 + bool resize (const Geometry &);
13
14 bool resize (int x, int y, int width, int height,
15 int border = 0);

const correctness

31 int x1, x2, y1, y2;
32
33 - CompWindow::Geometry geom = priv->window->geometry ();
34 - CompWindowExtents output = priv->window->output ();
35 + const CompWindow::Geometry &geom = priv->window->serverGeometry ();
36 + const CompWindowExtents &output = priv->window->output ();

more const correctness

35 + const CompWindow::Geometry &geom = priv->window->serverGeometry ();
36 + const CompWindowExtents &output = priv->window->output ();

Lots of this - in paint code we want to use serverFoo because it was the last thing sent to the server. And this is always guarunteed to come back to us.

166 - unsigned int width = window->size ().width ();
167 - unsigned int height = window->size ().height ();
168 + unsigned int width = window->geometry ().width ();
169 + unsigned int height = window->geometry ().height ();

CompWindow::size was just returning CompSize (priv->width, priv->height) which was just redundant data which should be removed. Replaced it with geometry () for clarity

299 + void move (int dx, int dy, bool sync);
300 + bool resize (int dx, int dy, int dwidth, int dheight, int dborder);
301 + bool resize (const CompWindow::Geometry &g);
302 + bool resize (const XWindowAttributes &attrib);
303 +

Added PrivateWindow::move and PrivateWindow::resize - to be called whenever a new position / size comes back from the server. (cut'n'paste code from CompWindow::move and CompWindow::resize

335 -
336 - gettimeofday (&lastConfigureRequest, NULL);
337 - /* Flush any changes made to serverFrameGeometry or serverGeometry to the server
338 - * since there is a race condition where geometries will go out-of-sync with
339 - * window movement */
340 -
341 - window->syncPosition ();
342 -
343 - if (serverInput.left || serverInput.right || serverInput.top || serverInput.bottom)
344 - {
345 - int bw = serverGeometry.border () * 2;
346 -
347 - xwc.x = serverGeometry.x () - serverInput.left;
348 - xwc.y = serverGeometry.y () - serverInput.top;
349 - xwc.width = serverGeometry.width () + serverInput.left + serverInput.right + bw;
350 - if (shaded)
351 - xwc.height = serverInput.top + serverInput.bottom + bw;
352 - else
353 - xwc.height = serverGeometry.height () + serverInput.top + serverInput.bottom + bw;
354 -
355 - if (shaded)
356 - height = serverInput.top + serverInput.bottom;
357 -
358 - if (serverFrameGeometry.x () == xwc.x)
359 - valueMask &= ~(CWX);
360 - else
361 - serverFrameGeometry.setX (xwc.x);
362 -
*snip*
616 - XMoveResizeWindow (screen->dpy (), id, 0, 0,
617 - serverGeometry.width (), serverGeometry.height ());
618 - window->sendConfigureNotify ();
619 - window->windowNotify (CompWindowNotifyFrameUpdate...

824 -
825 - if (!pendingConfigures.pending ())
826 - {
827 - /* Tell plugins its ok to start doing stupid things again but
828 - * obviously FIXME */
829 - CompOption::Vector options;
830 - CompOption::Value v;
831 -
832 - options.push_back (CompOption ("window", CompOption::TypeInt));
833 - v.set ((int) id);
834 - options.back ().set (v);
835 - options.push_back (CompOption ("active", CompOption::TypeInt));
836 - v.set ((int) 0);
837 - options.back ().set (v);
838 -
839 - /* Notify other plugins that it is unsafe to change geometry or serverGeometry
840 - * FIXME: That API should not be accessible to plugins, this is a hack to avoid
841 - * breaking ABI */
842 -
843 - screen->handleCompizEvent ("core", "lock_position", options);
844 - }

DIE

924 -bool
925 -PrivateWindow::checkClear ()
926 -{
927 - if (pendingConfigures.pending ())
928 - {
929 - /* FIXME: This is a hack to avoid performance regressions
930 - * and must be removed in 0.9.6 */
931 - compLogMessage ("core", CompLogLevelWarn, "failed to receive ConfigureNotify event on 0x%x\n",
932 - id);
933 - pendingConfigures.dump ();
934 - pendingConfigures.clear ();
935 - }
936 -
937 - return false;
938 -}
939 -
940 void
941 compiz::X11::PendingEventQueue::add (PendingEvent::Ptr p)
942 {
943 @@ -2454,21 +2148,6 @@
944 mValueMask (valueMask),
945 mXwc (*xwc)
946 {
947 - CompOption::Vector options;
948 - CompOption::Value v;
949 -
950 - options.push_back (CompOption ("window", CompOption::TypeInt));
951 - v.set ((int) w);
952 - options.back ().set (v);
953 - options.push_back (CompOption ("active", CompOption::TypeInt));
954 - v.set ((int) 1);
955 - options.back ().set (v);
956 -
957 - /* Notify other plugins that it is unsafe to change geometry or serverGeometry
958 - * FIXME: That API should not be accessible to plugins, this is a hack to avoid
959 - * breaking ABI */
960 -
961 - screen->handleCompizEvent ("core", "lock_position", options);
962 }

DIE

GOOD IN THEORY:

* Most of the description sounds like steps in the right direction.

* Simplified: Replacing 559 lines of code with 247.

UNSURE ABOUT THE THEORY:

* "composite and opengl now use the geometry last sent to the server"
This should give maximum performance, sure, but what if (for some reason) the server rejects the configure request as being somehow invalid? Would you ever encounter that, like moving a window to an invalid location?
For stability and correctness (which should be the primary goal right now) don't we want "composite and opengl now use the geometry last received from the server"?

* Caveat #1 is not really a caveat. It's the right (intended) design. But if you want to see a (buggy) example of how to solve the chattiness look at the timer code here:
https://code.launchpad.net/~vanvugt/compiz-core/fix-764330-trunk/+merge/86497
This was slightly buggy though as it broke keyboatrd-initiated moves (incorrect pointer warp). Simple to fix, probably.

* Still no XFlush's to ensure XConfigureWindow happens immediately. Or are there adequate XSyncs already?

BAD IN THEORY:

* The diff is too big for a mere mortal (who hasn't worked on all the code before) to fully load into their mental interpreter and evaluate its theoretical correctness. Please read that sentence again; it's important.

GOOD IN PRACTICE:

* Window movement is smooth and predictable, but not really better than in oneiric.

BAD IN PRACTICE:

* The bug introduced is worse than the bug being fixed; Window contents offset from the frame/decorations whenever a window is created or maximized.

* All that code, but we still haven't reduced the lag below oneiric-levels. Still not as good as Gnome Shell.

CONCLUSION:

The theory is mostly good but there is at least one serious bug that needs fixing. We probably shouldn't risk 0.9.7.0 any more with large changes like this. We should aim to get this code into a later release when it's more stable, not 0.9.7.0.

> GOOD IN THEORY:
>
> * Most of the description sounds like steps in the right direction.
>
> * Simplified: Replacing 559 lines of code with 247.
>
>
> UNSURE ABOUT THE THEORY:
>
> * "composite and opengl now use the geometry last sent to the server"
> This should give maximum performance, sure, but what if (for some reason) the
> server rejects the configure request as being somehow invalid? Would you ever
> encounter that, like moving a window to an invalid location?
> For stability and correctness (which should be the primary goal right now)
> don't we want "composite and opengl now use the geometry last received from
> the server"?

See http://tronche.com/gui/x/xlib/window/XConfigureWindow.html

Since we have SubstructureRedirectMask, it is guarunteed that any ConfigureWindow request we make on a child of the root window is guarunteed to end up in that position (hence the event tracker).

The only error cases I can think of is where you ask for a window to be of a size < 0 or > MAXSHORT. Ideally we should check for that case.

>
> * Caveat #1 is not really a caveat. It's the right (intended) design. But if
> you want to see a (buggy) example of how to solve the chattiness look at the
> timer code here:
> https://code.launchpad.net/~vanvugt/compiz-core/fix-764330-trunk/+merge/86497
> This was slightly buggy though as it broke keyboatrd-initiated moves
> (incorrect pointer warp). Simple to fix, probably.
>

There are some optimizations that can be done here - for example, if a window is grabbed, you can withold sending positions to the server until it is ungrabbed, or there is some other reason why you would want to do that. When that happens you can send it all at once.

> * Still no XFlush's to ensure XConfigureWindow happens immediately. Or are
> there adequate XSyncs already?

The GLib mainloop work introduced an XFlush as soon as we have finished dispatching the X event source.

>
>
> BAD IN THEORY:
>
> * The diff is too big for a mere mortal (who hasn't worked on all the code
> before) to fully load into their mental interpreter and evaluate its
> theoretical correctness. Please read that sentence again; it's important.

Indeed. The good thing is that its mostly all deletion and consolidation.

>
>
> GOOD IN PRACTICE:
>
> * Window movement is smooth and predictable, but not really better than in
> oneiric.
>
>
> BAD IN PRACTICE:
>
> * The bug introduced is worse than the bug being fixed; Window contents offset
> from the frame/decorations whenever a window is created or maximized.

Alright. I haven't seen this yet, but I have an idea of what might cause it. Will have a look into it when I get some time.

>
> * All that code, but we still haven't reduced the lag below oneiric-levels.
> Still not as good as Gnome Shell.

This will be fixed once the chattyness goes away. (as above). Watch your X.org CPU usage as you move windows - it skyrockets because at the moment we flood it with requests.

>
>
> CONCLUSION:
>
> The theory is mostly good but there is at least one serious bug that needs
> fixing. We probably shouldn't risk 0.9.7.0 any more with large changes like
> this. We should aim to get this code into a later r...

OK, I'm still wading through the changes. But I feel like a rant...

1307 CompSize
1308 CompWindow::size () const
1309 {
1310 - return CompSize (priv->width + priv->geometry.border () * 2,
1311 - priv->height + priv->geometry.border () * 2);
1312 + return CompSize (priv->geometry.width () + priv->geometry.border () * 2,
1313 + priv->geometry.height () + priv->geometry.border () * 2);
1314 }

As a change this doesn't look too bad - but it ignores a horrid design!

1. Chained operations like "priv->geometry.width ()" imply too muck knowledge of the details of "priv".
   That is "priv->width ()" would reflect less coupling.

2. You're tracing the path "priv->geometry..." many times, which suggests that the logic belongs in "geometry".
   "return priv->geometry.size ()" or "return CompSize(priv->geometry)" would reflect better coherence.

So, assuming (because borders may be optional?) that there's an unambiguous mapping from CompWindow::Geometry:

a. Add a constructor to CompSize: "explicit CompSize(CompWindow::Geometry const& g) : mWidth(g.width () + g.border () * 2) ..."
b. Add an inline method "CompSize PrivateWindow::size () const { return CompSize(priv->geometry); }"
c. Rewrite the above as "CompSize CompWindow::size () const { "return priv->size ()"; }"

264 - /* been shaded */
265 + /* been shaded
266 if (w->priv->height == 0)
267 {
268 if (w->id () == priv->activeWindow)
269 w->moveInputFocusTo ();
270 - }
271 + }*/

Comments are not for storing old versions of the code - that's what we use bzr for. ;)

There's a lot of "<blah>.width () + <blah>.border () * 2" and "<blah>.height () + <blah>.border () * 2" around. Surely CompWindow::Geometry could have a "widthWithBorders" method - or maybe a free function?

I'm tempted by

template<Typename T>
inline int heightWithBorders(T const& blah) { return blah.height () + blah.border () * 2; }

> OK, I'm still wading through the changes. But I feel like a rant...
>
> 1307 CompSize
> 1308 CompWindow::size () const
> 1309 {
> 1310 - return CompSize (priv->width + priv->geometry.border () * 2,
> 1311 - priv->height + priv->geometry.border () * 2);
> 1312 + return CompSize (priv->geometry.width () + priv->geometry.border ()
> * 2,
> 1313 + priv->geometry.height () + priv->geometry.border () * 2);
> 1314 }
>
> As a change this doesn't look too bad - but it ignores a horrid design!
>
> 1. Chained operations like "priv->geometry.width ()" imply too muck knowledge
> of the details of "priv".
> That is "priv->width ()" would reflect less coupling.

Probably, although priv is just a private member with implementation details, I don't really see this as a large design problem.

>
> 2. You're tracing the path "priv->geometry..." many times, which suggests that
> the logic belongs in "geometry".
> "return priv->geometry.size ()" or "return CompSize(priv->geometry)" would
> reflect better coherence.
>
> So, assuming (because borders may be optional?) that there's an unambiguous
> mapping from CompWindow::Geometry:
>
> a. Add a constructor to CompSize: "explicit CompSize(CompWindow::Geometry
> const& g) : mWidth(g.width () + g.border () * 2) ..."
> b. Add an inline method "CompSize PrivateWindow::size () const { return
> CompSize(priv->geometry); }"
> c. Rewrite the above as "CompSize CompWindow::size () const { "return
> priv->size ()"; }"

+1 for all three

> 264 - /* been shaded */
> 265 + /* been shaded
> 266 if (w->priv->height == 0)
> 267 {
> 268 if (w->id () == priv->activeWindow)
> 269 w->moveInputFocusTo ();
> 270 - }
> 271 + }*/
>
> Comments are not for storing old versions of the code - that's what we use bzr
> for. ;)

Indeed, this is a small portion of the code and I'm not yet sure what to do with it. When I get around to revisiting this section (when I actually get time to look at this again, wanted to get it up for review early) I'll see what can be done.

> There's a lot of "<blah>.width () + <blah>.border () * 2" and "<blah>.height
> () + <blah>.border () * 2" around. Surely CompWindow::Geometry could have a
> "widthWithBorders" method - or maybe a free function?
>
> I'm tempted by
>
> template<Typename T>
> inline int heightWithBorders(T const& blah) { return blah.height () +
> blah.border () * 2; }

I'm not really sure ?

The complexity here comes from a legacy part of the X Server coming into play - a window could have fixed dimentions, but also specify a "border" which would be exclusive of the fixed dimentions, so you'd have to take this into account for all positioning operations (Xterm comes to mind here).

I definitely agree that geom.width () + geom.border () * 2 feels fragile and indeed, that has tripped me up many times before. Maybe a default parameter with a bitflag makes sense here, eg IncludeBorder , IncludeBorderFirst , IncludeBorderSecond (as it is on both sides)

I'll note that the above isn't *really* all that relevant to this merge though, but good things to keep in mind in any case. I'd rather not see the diff get any better except for adding tests.

Please resubmit for target branch lp:compiz-core (0.9.7)