I managed to get my hands on the N900 after a long long wait. But the path to getting the device was long and tragic. Here’s the story.

Back at foss.in (Dec 1st week 2009), I heard of a contest held by the Nokia maemo-bangalore team.  The sexy 3d pic on the blog got me all excited. So, at foss.in, including all my customer commitments, I worked 3 days/nights at a stretch to write a youtube video browser, cilantro (anagram of the irc nick of a good friend and lead of qt/webkit team to whom the app is dedicated to ).  In all honestly, I put myself under some pressure to participate in the contest – I was already overloaded with work but I thought that if I can get my hands on the device I can improve QtWebKit Flash performance  since I was working on that at that point. So, you can understand my excitement when awaiting for the results.

Except it was not announced (we were told informally that it will be announced at end of foss.in). We were told it was postponed to be on the 8 Dec 2009. Minor disappointment. But, fine. That didn’t happen either. After a few tweets and poking a nokia friend, the results were announced on 11 Dec 2009. No mail to winners, no notification on when we will get the devices. Just the blog.  I mean no “thanks for participating, we will announce winners on xx date” even. Oh well. I forgive, only because I won. But my joy wasn’t long lived.

To my shock, the winners had folks who had not written their app for the contest. Atleast following the links on the blog suggests that the apps were published way back. Surely, one expects code to be written around the time of the contest? I must be wrong, so to make a more judicious decision, I went around looking for the source. Except, with all my googling powers, I cannot find the source for any (except one which had a commit long before contest announcement). Still convinced that my assumptions were just wrong, I left a comment on the blog asking for the code. No response. I sent a mail asking for the code. No response. So, I poked the Nokia friend yet again. This time I got a response saying they *cannot* publish the code and they will send a mail to just let people upload code to maemo garage if they care. HUH!? So, here we have a contest for foss.in where the winners have been adjudged with no published code. Sad, really sad.

On dec 18, I was told it *might* take a month for the device to arrive (and only after I asked them). 1 month!? I mean I wrote this app to get my hands on it immediately. Maybe I am expecting too much. I wait. On Jan 8, I asked them what the status was. No response (see the trend?). So I poked Nokia friend again and I got a response on Jan 14, saying they will intimate me by phone. As, you can guess, nothing was heard from them ever again. On Feb 3, I mailed them again. No response as usual. This time I didn’t ask my Nokia friend. I had given up on the people running the contest and I was going to take this outrageous behavior public.

Amazingly on Feb 16 2009, I got contacted by the maemo team for the first time ever (i.e first time initiated by them). And I did end up getting the N900 but the whole thing was such a sorry experience that I will never enter a contest that involves the maemo-bangalore team. I am not even sure if I deserve the award and if the contest was a FOSS contest (we can never know until we see the source). I had given them the link to my source but they haven’t bothered updating the blog post.

@maemo-bangalore – you need to learn to communicate and in the least, atleast respond when asked questions. You should be ashamed of treating the community and contest winners this way. If you don’t have resources to do a good job, just don’t run contests. What you have done is a mockery of FOSS. I am not actually upset with the late delivery of the device but the way you ran the contest.

If I were Nokia, I would fire the incompetent person who ran the contest.

Over the past couple of months, I have been working losing sleep over getting Flash to work in QGraphicsView on all platforms.  The good news is that flash now works in QGV on all platforms. Yes, that means that not only can QGV display Flash but can also rotate and transform flash when using QGraphicsWebView. IMPORTANT: Flash embedded in a QWebView using QGraphicsProxyWidget is NOT supported.

What follows is a brief summary of my work and the list of patches that you will need to get Flash working with Qt 4.6.0. Note that when I say plugin below, I mean NPAPI plugins like Flash. They have nothing to do with QWidget based plugins. QWidget based plugins should work in Qt 4.6.0 when using QGraphicsWebView. I have also simplified the concepts (in favor of being completely correct) in the post below so that it is readable.

Windowed and Windowless plugins

Flash gets embedded into web pages using an API called NPAPI (Netscape Plugin API). NPAPI provides two mechanisms to embed plugins into a web page – Windowed mode and Windowless mode.

Windowed mode – the idea is that the plugin gets a native widget handle and does it’s own mouse, keyboard processing. For example, on Windows, we give the plugin a HWND. The plugin will install a window proc and process events directly (i.e the events are never seen by the browser). On X11, the plugin embeds itself into the Window using XEmbed. When the browser scrolls, the browser moves the native window appropriately. Mac doesn’t support windowed mode.

Windowed mode should have been enough but it has a limitation that it interferes with the z-ordering of html elements.  Some sites like to put html _over_ their flash content (for example, menus). This can be achieved by making the menu html have a higher z-order than the plugin. Can you spot the conflict interest here between z-ordering of native widgets and z-ordering of HTML elements? The plugin native window is a child of the browser’s window and we have a situation where the browser wants to paint HTML after the child has painted itself! This is not possible since a parent window always draw before it’s child. There are other disadvantages too – it is also impossible to transform plugins (native widgets are rectangular), making them transparent is tricky and printing plugins requires lots of hacks.

Windowless mode – The idea here is that plugins just paint to a pixmap – platform’s graphics context to be more precise (HDC on Windows, Pixmap on X11, ContextRef on Mac). It is now the browser’s responsibility to pass (forward) mouse and keyboard events to the plugin. Since the browser now contains a screenshot/snapshot/pixmap of the plugin, the browser can transform plugins, draw it in the correct z-order, make is transparent even. The main disadvantage here is that it can tend to be a bit slow.

Evan has a better explanation of the modes. There’s also one by Robert from which I originally learnt about these modes but I cannot find the link anymore).

Flash

I keep saying Plugin but really we care only about Flash At least, for the moment that is our priority. All version of Flash supports Windowed mode. Only Flash 10 supports windowless mode on linux.

Flash lets the html/plugin author control the mode of operation using the wmode parameter (i.e pass it as part of html <object> tag). By default, it operates in windowed mode. Setting wmode to ‘transparent’  makes the html below visible through the flash. If wmode is set it to ‘opaque’, the html will not be seen through and the flash will have some solid background color (think of it as Qt::WA_OpaquePaintEvent, it allows for optimizations). transparent and opaque mode are implemented using ‘windowless’ mode. Here’s a great site to test out these modes in action.

Important: When in windowed mode, Flash expects a native window handle/id. A browser cannot force windowless mode – the mode is decided by the plugin. A browser can, in theory, say that it supports only windowless mode but this makes Flash crash (at least) on Linux. Hence, when we have a site like youtube (which uses Flash in windowed mode) in QGV, we need to come up with a way of providing flash with a window handle. There are two ways to deal with this problem – we can either give Flash a fake native window, we then grab the contents of this fake native window and paint it wherever we want OR we can inject the wmode ‘opaque’ parameter when loading flash in QGV. The former approach is platform specific and requires lots of work. The latter is a 10 line patch.

Plugins in QGraphicsView

Items inside QGraphicsView are not native widgets. They are just items on a canvas that can be transformed and visualized using multiple views. This means that windowed mode of plugins is not feasible (apart from the fact that we have no native widget handle for items on canvas). To make plugins work in QGV, we need to implement windowless mode.

X11 (Linux)

Windowed mode on Linux has already been implemented in Qt 4.5. What was missing is the windowless mode support. One nice feature is that Flash is able to fall back windowed mode when the browser says it cannot support windowless mode (vice versa crashes as I noted before). This is the reason that sites that use windowless mode work in QWebView with Qt 4.5.

To support Flash in QGV, I started out adding windowless mode support. In theory, implementing windowless mode is a matter of passing a X Pixmap to the plugin, but there were lots of problems. The entire history is on bugzilla bug 20081.

• Flash uses a different X Display connection – the default one provided by gtk/gdk and not the one that Qt provides it. Ouch. This means that when flash paints something, we have to XSync to get changes reflected in the Qt connection and vice versa. The uglier part is that Qt somehow needs to get hold of gdk’s Display without linking to gdk. It’s all very hairy. (r49158).
• Flash uses a different X Visual (the default system visual) rather than the one provided by Qt. This means that even if we created a 32-bit pixmap, Flash won’t be able to draw into it because it uses a 24-bit visual. Since people really wanted transparency to work, the solution is to grab the contents of the backing store to fake transparency. Very hairy stuff. This solution works in QWebView but not with QGraphicsWebView (r49169). We have turned off transparency in QGVuntil this Flash bug is fixed.
• The NPAPI requires X Pixmap. QPixmap may or may not be backed by a X Pixmap depending on thegraphicssystem.
• Print preview implementation of Qt holds a reference to the X Pixmap (because of QPicture). This meant that when the user prints we have to ‘let go’ of the X Pixmap because the print preview has to remain unaffected by on screen changes. At the same time, someone needs to destroy this X Pixmap when the print preview dialog close. This one was very tricky to solve. Actually, I didn’t solve it, Samuel did (r50123)

All windowless mode changes for Linux are part of Qt/WebKit 4.6. QGV on linux can display Flash only in windowless mode. If Flash is in windowed mode (like in youtube), it does not work.  As mentioned there are two methods to fix windowed mode in QGV . Method 1 – we create a fake window and grab it’s contents. This can be achieved using X Composite and X Damage (as done in Fennec). I have a patch at 31232 but it requires more work. Method 2 – we can inject wmode opaque. This has been committed as part of 32059. If you need youtube to work in QGV, you need to apply that the patch r51759 on top of Qt 4.6.0.

Mac OS X

On the Mac, plugins operate in Windowless mode _always_.  There is no such thing as windowed mode on the Mac. NPAPI  on Mac supports various event models and painting models. Event model dictates how we pass the mouse/keyboard events to Flash (i.e which data structure is used to report events – NSEvent or the classic EventRecord).  Qt/WebKit only supports only the carbon event model and the CoreGraphics drawing model (there’s a open gl drawing model, core animation and a QuickDraw drawing model). ATM, only Carbon based apps can display flash. Qt/Cocoa apps (32 or 64 bit) cannot display Flash in QWebView or QGV. (Confusingly, this has nothing do with Qt/WebKit not supporting cocoa event model. It is possible for Qt/Cocoa apps to display using the carbon event model. This is tracked using 32376)

Now to the carbon event model – how it works is that the browser provides the plugin it’s window handle (a WindowRef), a ContextRef and the position (a rectangle) inside the window and Flash draws itself there.

Simple, no? You wish Unlike in QWebView, one cannot provide a rectangle as the position of Flash since the Flash could be transformed. For those new to Mac, ContextRef is like a QPainter (it can be over any paint device – a widget, pixmap, image). What we do is to get the ContextRef of a QPixmap and pass the position as (0, 0, width, height). Unfortunately, Flash requires the WindowRef parameter to be valid because it seems to be using it to detect if the window is ‘active’ for processing mouse move events. To workaround, we create a fake hidden window. In addition, we also fake the fact that this fake hidden window is the active window (else Flash won’t process mouse move events)! If you thought this was insane, you should check out what the chrome devs do (they intercept carbon calls of Flash by injecting a library using DYLD_INSERT_LIBRARY at runtime). See 31183 , 31794, 31979 for more history.

None of the patches are in Qt 4.6.0. You need to apply  r51234 for painting; r51105, r51412 for mouse handling andr51485 for context menu position. Currently, printing does not work on the Mac. This is tracked at 31975. Flash transparency works fine in QGV.

Windows

Windows has always supported Windowed and Windowless mode (Thanks to Apple). Printing already works thanks to some awesome hacks. Flash transparency works in QGV. However, the problem (as mentioned above) is that Flash expects a HWND when in windowed mode. So, youtube inside QGV requires a HWND. Fix is to inject wmode opaque. You need to apply r51979 on Qt 4.6.0

Symbian

The hardwork here was done by Yael. See 29302 for more details.

Gitorious

I have done the necessary cherry-picking and published a branch (based of 4.6.0) athttp://qt.gitorious.org/~girish/qt/girishs-qt/commits/4.6.0-gv-flash. With that branch, Qt should support Flash in QGV on all platforms in all modes.

More information

If you need more information on internal working of plugins, Yael, Torarne and I have writtenhttp://trac.webkit.org/wiki/QtWebKitPlugins. Drop by on irc channel #qtwebkit on freenode for any clarification. Performance needs some love, we are working on it.

My work would not have been possible without the tireless reviews of Simon, Holger, Kenneth. Thanks guys! Simon, in particular, who  kept pushing me to try out various possibilities/hacks. I cannot thank him enough! And it’s a matter of great pride that I am now a webkit commiter!

The guys at Linden Lab have made ubrowser support Qt/WebKit (llqtwebkit). With the above changes, flash works in ubrowser.

Almost a year back, I was looking for a maintainer for kdocker. I had practically given up, when out of the blue John contacted me to be the new maintainer. What more, he had already ported it Qt4! That made handing over maintainership to him a no-brainer

Also, KDocker is now moved from sourceforge to launchpad. Once I regain access to sourceforge, I will update the sourceforge page.

Read the release announcement here. If you are interested in helping  out with translating the app, please contact John.

Integrating the web into desktop apps is now incredibly easy thanks to Qt WebKit.  Qt WebKit also allows us to embed any Qt widget inside the html. This feature makes it extremely tempting to develop full blown desktop applications using Qt WebKit – your interface is entirely designed using html/css (instead of .ui) and you as a C++ developer decide what parts are best done in Qt/C++ and embed them into the html page.

The above scheme works very well (for the most part). However, there are a lot of features that make our applications feel native, that we take for granted in desktop apps when using Qt/C++ but not so when using Qt WebKit. So here is a list of short comings that you should know about.

1. Focus handling - Tab focus  doesn’t work at all when navigating from html into a c++ plugin and vice versa.

2. Keyboard accelerators - In html, you can specify accelerator using ‘accesskey’. They don’t seem to work in QtWebKit. They do work in google chrome, so must be something to do with the Qt port.

3. Text can be selected and images can be dragged – Your main ui is html, what did you expect ?

4. Plugins will get deleted if you hide them – If you hide your C++ plugin/widget using ‘display: none’, Qt/WebKit will delete it. In your traditional Qt program, all your data is part of the widget – now it’s all gone. To prevent this, you have to write all your widgets to use the model-view paradigm. An alternative is to use ‘visibility: hidden’ which won’t delete the plugin but the plugin will continue to hog space in the layout. Side note: In Qt 4.4.x, plugins weren’t deleted (aka leak). Starting Qt 4.5.x, they were deleted – we found this problem when our app starting crashing gloriously when we moved to Qt 4.5.x.

5. Fluid layouts - When the windows resizes, UI elements inside are expected to reposition and resize nicely; the window must have a nice minimum size – that is all done by QLayout does. This is really hard to get right when using html. You will most likely end up writing js layouting code.

6. Moving data back and forth is slow – Passing Qt types from C++ to JS works very well using QJson. Converting complex types like a QPixmap to html img is really really slow, since you have to convert in into png first. Also, using this method crashes Asus laptops/Windows, probably some graphics card issue, nevertheless your app has to deal with it.

8. RTL – Qt will layout your widgets automatically when your app is run in a RTL environment. With HTML, you will have to provide an alternate html file, do your own RTL detection, implement sizing fixes to elements for different languages and so on.

9. Translations – Qt provides you linguist, lrelease, lupdate. When using HTML, you will have to roll out your own translation framework (this is an awful lot of work).

10. Misc – There are other minor problems depending on your Qt version - the default fonts of html and qt didn’t match. There was a time when a 12pt font of Qt and 12 pt font of webkit don’t match (this seems to be fixed now). Displaying plugins inside divs had some artifacts.

Note that the above is not a case against using Qt/WebKit for developing desktop apps since they can all be worked around. It’s more for you to know what you will be up against if you decide to go down this road.

Update 1: Added note on rtl and translations