Access to raw sockets in Android instrumented tests

Question

I've trawled through many of the answers here regarding socket access via native code on Android, and others regarding socket access for "androidTest" instrumented tests, but none are able to fully explain some odd behaviour I'm seeing.

Android Studio 2.3.3 on Windows 10, NDK 15.1.x, build tools 25.0.3. Building with Cmake. I am porting some native code over to Android intended to be distributed as a library that is wrapped in a Java JNI-based API. This part seems to be working fine; at least I can debug and log into the native code and see where things go wrong.

I've created a few instrumented tests to exercise the JNI, but it looks like the native side does not have access to sockets, even if the test app that Android Studio wraps up for your intrumented tests does (i.e., it has the INTERNET permission applied, and I can see that it is part of the pushed manifest of the test app. I also applied the ACCESS_NETWORK_STATE perm in a fit of pique.)

That is, parts of the library know how to set up and use TCP sockets (datagram and stream; in this case stream) and resolve DNS, which fails (This might be a bug in my port, since the device itself still seems to resolve DNS based on the adb shell output below). If I test with an IP address it retries until it fails. Each call to socket() returns an ERRNO of 11 EAGAIN ("Try again").

If I use the adb shell to login to the device under test over USB, I can ping and use curl, etc. But, as soon as I run-as as the test app I am no longer allowed to use any network device.

shell@venice:/data/data/org.clvrmnky.library.test $ ping www.example.com
PING www.example.com (93.184.216.34) 56(84) bytes of data.
ping: sendmsg: Operation not permitted
ping: sendmsg: Operation not permitted
ping: sendmsg: Operation not permitted
^C
--- www.example.com ping statistics ---
3 packets transmitted, 0 received, 100% packet loss, time 2004ms

1|shell@venice:/data/data/org.clvrmnky.library.test $ ping -I wlan0 www.example.com
ping: SO_BINDTODEVICE: Operation not permitted
2|shell@venice:/data/data/org.clvrmnky.library.test $ curl -4 --verbose www.example.com/
*   Trying 93.184.216.34...
* connect to 93.184.216.34 port 80 failed: Connection timed out
* Failed to connect to www.example.com port 80: Connection timed out
* Closing connection 0
curl: (7) Failed to connect to www.example.com port 80: Connection timed out
7|shell@venice:/data/data/org.clvrmnky.library.test $

I won't paste it here, but wlan0 (and others) exists and is UP with a valid, reachable IP address. It is "Link encap:UNSPEC" which I admit I don't fully grok.

I'm making an assumption that the app failing on the successive socket calls and the inability to bind to a network device in and use it in an adb shell are related in some manner, but if someone has a reason why this may not be so, please let me know.

I've tried:

• Invoking the test methods in a background thread, just in case the instrumented tests were running in the test app main thread.
• Setting the StrictMode thread policy to permitAll() in the @Before clause of the test class.
• Running with and without the INTERNET permission set in the test manifest. Without this permission it reaches the native code, but fails much earlier in the initialization. I'd have to dig into my notes to fetch more details about that. I assume that I need this permission. (Side note: if I enable INTERNET then the remote debugger can no longer attach to the device, so I can debug via logs only.)

Answer 1

It looks like instrumented tests, at least as of this writing, do not support access to raw sockets through the NDK. The .test app that is auto-created seemed to have everything it needed in the manifest (I checked base.apk right from the device).

Once a DemoApp was created with identical manifest settings I was able to open sockets, connect to them, do name lookups, etc. Furthermore, I was able to access the network via adb shell when running as ("run-as") the DemoApp.

There is still something strange going on, as I created a separate JNI app that opens sockets, and hacked an instrumented test against that. At least briefly adb shell running as ext.other.app.test did have internet access.