Initially, the scanner used libmagic which used magic numbers in the file's
content to detect what kind of file it appears to be. Since that library isn't
available on all systems, we added support for two other libraries, mimetypes
amongst them.
The issue with mimetypes is that it only uses the file's extension, not its
actual content. So this ends in variable behaviour depending on what system
you're using fdroidserver on. For example, an executable binary without
extension would be ignored if mimetypes was being used.
We now drop all libraries - mimetypes too as it depends on the system's
mime.types file - and instead check extensions ourselves. On top of that, do
a simple binary content check to find binary executables that don't have an
extension.
The new in-house code without any dependencies doesn't add any new checks, so
no builds should break. The current checks still work:
% fdroid scanner app.openconnect:1029
[...]
Found executable binary at assets/raw/armeabi/curl
Found executable binary at assets/raw/mips/curl
Found executable binary at assets/raw/x86/curl
Found JAR file at lib/XposedBridgeApi-54.jar
Found JAR file at libs/acra-4.5.0.jar
Found JAR file at libs/openconnect-wrapper.jar
Found JAR file at libs/stoken-wrapper.jar
Found shared library at libs/armeabi/libopenconnect.so
Found shared library at libs/armeabi/libstoken.so
Found shared library at libs/mips/libopenconnect.so
Found shared library at libs/mips/libstoken.so
Found shared library at libs/x86/libopenconnect.so
Found shared library at libs/x86/libstoken.so
Advantages:
* Easier to tell why we need each package
* apt-get install output is less scary/huge
* CI job is split in more, smaller steps easier to debug
the `android` utility is pretty stupid, it doesn't really cache the
package index info. So each time it is run, it tries to fetch the
indexes from the network. This combines all android package installs
to a single command to make things run quicker.
