Before I start: I am going to say “Flash” in this blog post a few time. I am also going to say “Fennec”. I am not going to say “I like Flash” or “I do not like Flash”. I am also not going to say “We are shipping Flash support”. This blog post is about the progress we have been making on getting NPAPI and Flash working in Fennec. Decisions to ship such support is another topic. Please think of this as a feasibility study… an informal one.
I went for a short vacation and recently picked up the Android NPAPI implementation for Fennec. When I left, we had some basic bitmap drawing going. As you may recall, the Webkit Android NPAPI has a few different drawing models. The first one (and easiest to understand and get bootstrapped) is called Bitmap Model. It basically is a piece of memory that you can hand to the plugin and the plugin will call back on a set of functions that you have registered with it. These sets of functions (interfaces) are a simple wrapper around Skia. So, we got the sample plugin loaded and saw a blue ball bouncing around. Yay.
Next step was to get the next drawing model working. This one is called the Java Surface Model. Two things about this drawing model. First, it is super important as it is the model that we believe that Adobe Flash uses exclusively. If we get this model right, the other’s are far less important. The other thing is that this model is very scary as it uses Java. Recall that Fennec doesn’t have access to the Java runtime in the child process. (because for responsiveness Fennec uses multiple processes — one for the UI, one for the web content. The web content process doesn’t have access to Java). So, while I was on the beach and making tough decisions like which beach to go to tomorrow, Brad Lassey started reflecting the JNI interface over our IPC layer. Yes, he is fearless. His work allows you to make JNI calls from the child. So, if you want to create a new Java string, you can use his interface from the child process and it will return an opaque pointer to a new Java string that was created in the parent. This is not meant to be a general JNI remoting service (but could develop into one). The intent is to only get as much JNI remoting needed in order to load Flash.
So, that brings us to my last hack session. The way this drawing model works is that the plugin asks you what you support. We say, we support these drawing models, Bitmap and Java Surface. The plugin we are targeting rather do things with a Java Surface, so it asks that we use that drawing model. After the plugin finishes draw model negotiation and initialization, it is ready to go. We just have to get its “surface” and render it to the screen. To do this, we make a standard NPAPI GetValue call to the plugin and use the define kJavaSurface_ANPGetValue. Here is the comment that is in the android_npapi.h header file:
/** Requests that the plugin return a java surface to be displayed. This will
only be used if the plugin has choosen the kSurface_ANPDrawingModel.
NPP_GetValue(inst, kJavaSurface_ANPGetValue, jobject surface)
#define kJavaSurface_ANPGetValue ((NPPVariable)2000)
A couple of sad points:
- This interface was never standardized and this GetValue may eventually collide with some another GetValue key. Not really sure what should be done here other than document it.
- The “surface” return isn’t a Surface at all. Instead, it is a View. The comment and the #define name are completely bogus. Looking at the sample plugin, it is clear the result of the GetValue call is a subclass of View, not of Surface.
Ignoring that, if the call is successful, the plugin should have created a View and will draw into it as soon as that View is attach to some View Manager. (Note: you can’t access the raw bitmap from this drawing model. If you could, it would integrate much nicer into Fennec.). To get this View into our ViewManager, we have to tell the parent process about it. This is basically where I ended last night. For a simple check point, I forwarded the View pointer to the parent, and absolutely placed it in our ViewManager. Guess what, it worked. We could see a little 3D object spinning around in the browser. Well, above the browser. Partial Yay.
Next steps are to clean the code up a bit, have the plugin object be position and scaled appropriately, and fix more bugs along the way.