AOSP/Bionic compatiblity layer for GLIBC systems
Currently heavy work in progress. Linker will be rewritten as it's
utter crap. Need to add logging/tracing facility controlled from env.
That's about it, otherwise it's just implementing stuff. libjvm-*.so
implementations are not correct at all, but the point is to not make
correct implementations yet, just to get some games run, which are pretty
forgiving for incorrect implementation :D.
# to compile for android ABI x86 <https://developer.android.com/ndk/guides/abis>
make x86Other than x86 is currently untested and probably doesn't work.
./andre my-android-binary
./andre my-android-jni-lib.so
sh andre-apk.sh my-android-unity-app.apkandre-apk.sh currently assumes app that uses libunity.so.
AOSP/bionic libs in runtime (the non libjvm-*.so ones) can be used to
port android software to GNU/linux also. The runtime libraries contain
bionic_ prefixed symbols for runtime compatibility only. Functions that
are not implemented in glibc, but are in AOSP/bionic are implemented
normally in these libraries. This means that the same libraries can be
used for both, compiling and linking to GNU/linux android native code, and
running AOSP/bionic linked code. The PORTING and RUNTIME capability is not
optional compile-time, nor do we separate PORTING and RUNTIME libraries.
This is for pure simplicity.
If you have to load/use bionic linked libraries from your ported
application, you can use the libdl.so that has bionic_ prefixed
versions of the usual dlfcn.h functions.
If for some reason one wants minimal library for either PORTING or RUNTIME task, here are the rules you can apply for symbol removal:
PORTING: Remove all symbols with prefix `bionic_`. Don't do this if you need `bionic_dlopen` from `runtime/libdl.so`.
RUNTIME: Remove the symbol if there also exists the same symbol with a `bionic_` prefix, retaining only the `bionic_` edition.
That said, the libraries are pretty small anyways.
We implement the JNI interface and simulate a rather unusual JVM that
only calls C functions. Native android code that uses JNI, thinks
it's talking to java but in reality it's calling functions from the
runtime/libjvm-*.so libraries. The runtime/libjvm-*.so libraries
implement java apis referenced by the application in C.
In future, there might be very small dalvik bytecode interpreter, that we can use to eliminate the need of porting internal java dependencies and implementing entrypoint for every native application. We would only have to implement the external java apis that exist in android system image in C, and yes, this also means you could run android java applications.
If you are porting android application to GNU/Linux that heavily relies
on JNI, you can use the libjvm.so to use the minimal JVM inside the
application, without having to rewrite the JNI reliant code. At this
moment however, you still have to manually do the initial JNI calls
that setup and initialize your application. This may not be needed
eventually, if we implemented what was described above (minimal dalvik
bytecode interpreter).
apkenv This is what inspired this project. The current linker code originates from this project (modified from AOSP). apkenv's in pretty messy state though and doesn't do clean implementation. It's still interesting approach and you should check it out. Also helpful resource for this project.
libhybris Very badly documented project, but I guess it has some overlap with this project? Seems to be mainly aimed for using android GPU drivers on GNU/Linux. Also uses AOSP linker, probably this is the resource apkenv used.
Currently it's not yet contribution friendly. When time comes, I'll write more detailed topic on this. Project also needs some regression testing for all the functions that will be implemented. As for implementing stuff, people should only implement the stuff some application needs. Lets not implement anything that doesn't get used.
Since apitrace and similiar OpenGL debuggers redirect the dlopen calls,
you have to LD_PRELOAD the runtime/libEGL.so while running apitrace.
runtime/libEGL.so is responsible for wrapping eglGetDisplay and
eglCreateWindowSurface functions, so that valid native handles are
passed to the real system EGL implementation.
LD_PRELOAD=runtime/libEGL.so apitrace -a egl ./app bioniclib.so
Basically, most android games are just native libraries with a very small java entrypoint that does some boilerplate app setup, and few calls to the native library. Native library may communicate back to java through JNI, but usually the java parts are very small in these kinds of applications. Thus these kind of applications will work fine with even barebones JNI/JVM implementation which doesn't neccessary even have to be very "correct" :). To use the native libraries on GNU systems though we need to use custom linker as bionic's linker is incompatible with glibc. Similarly bionic libc is incompatible with glibc and we have to handle these differences to be able to use android libraries on GNU/linux. We also have to implement the system libraries that exists on android, but not on traditional GNU/Linux, such as the stuff in AOSP that handle window/context creation, input and such. These implementations can be valuable for porters, or even application developers as some interfaces may be very clean.
While we could use dalvikvm to handle the JNI and java, we still would
have to reimplement all the android java apis to be GNU/Linux compatible,
or run android services in container or something. This is probably
what anbox and such projects do. The scope of this project is however
something that does not need anything from android system. The minimal
bytecode interpreter would be fun research project to see how little
code we can achieve all this with. Of course java's standard library
is quite large so the work for reimplementing java apis should not be
underestimated.
Our minimal JNI/JVM implementation is clearly separated from other AOSP/bionic components however, so it's perfectly possible for someone to use dalvikvm and android java libraries/services for the java bits instead. If someone's gonna do and maintain that work and it doesn't feel too clunky/hacky/invasive I have nothing against such code in this project. (aka, the JNI/JVM implementation is not hardcoded)
This is interesting and weird question at the same time. But yes, we could generate java classes that just specifies each method as native and thus calls our C implementation. However at current scope these implementations are not very useful. The goal is to just provide some barebones to get the interesting android applications to work. Correctness is not guaranteed! This project may be useful for NDK development however, basically anything where you don't use java much nor need a very "correct java implementation" :).
You can use qemu. For hw acceleration, we could use custom libGLES libs that communicate with host system. Though such libraries may exist already, maybe even inside qemu project.. I haven't researched.
