jnitrace

A Frida based tool to trace use of the JNI API in Android apps.

Native libraries contained within Android Apps often make use of the JNI API to utilize the Android Runtime. Tracking those calls through manual reverse engineering can be a slow and painful process. jnitrace works as a dynamic analysis tracing tool similar to frida-trace or strace but for the JNI.

Installation:

The easiest way to get running with jnitrace is to install using pip:

pip install jnitrace

Dependencies:

arm, arm64, x86, or x64 Android device
Frida installed on the Android device
Frida support > 12
Linux, Mac, or Windows Host with Python 3 and pip

Running:

After a pip install it is easy to run jnitrace:

jnitrace -l libnative-lib.so com.example.myapplication

jnitrace requires a minimum of two parameters to run a trace:

-l libnative-lib.so - is used to specify the libraries to trace. This argument can be used multiple times or * can be used to track all libraries. For example, -l libnative-lib.so -l libanother-lib.so or -l *.
com.example.myapplication - is the Android package to trace. This package must already be installed on the device.

Optional arguments are listed below:

-R <host>:<port> - is used to specify the network location of the remote Frida server. If a : is unspecified, localhost:27042 is used by deafult.
-m <spawn|attach> - is used to specify the Frida attach mechanism to use. It can either be spawn or attach. Spawn is the default and recommended option.
-b <fuzzy|accurate|none> - is used to control backtrace output. By default jnitrace will run the backtracer in accurate mode. This option can be changed to fuzzy mode or used to stop the backtrace by using the none option. See the Frida docs for an explanation on the differences.
-i <regex> - is used to specify the method names that should be traced. This can be helpful for reducing the noise in particularly large JNI apps. The option can be supplied multiple times. For example, -i Get -i RegisterNatives would include only JNI methods that contain Get or RegisterNatives in their name.
-e <regex> - is used to specify the method names that should be ignored in the trace. This can be helpful for reducing the noise in particularly large JNI apps. The option can be supplied multiple times. For example, -e ^Find -e GetEnv would exclude from the results all JNI method names that begin Find or contain GetEnv.
-I <string> - is used to specify the exports from a library that should be traced. This is useful for libraries where you only want to trace a small number of methods. The functions jnitrace considers exported are any functions that are directly callable from the Java side, as such, that includes methods bound using RegisterNatives. The option can be supplied multiple times. For example, -I stringFromJNI -I nativeMethod([B)V could be used to include an export from the library called Java_com_nativetest_MainActivity_stringFromJNI and a method bound using RegisterNames with the signature of nativeMethod([B)V.
-E <string> is used to specify the exports from a library that should not be traced. This is useful for libraries where you have a group of busy native calls that you want to ignore. The functions jnitrace considers exported are any functions that are directly callable from the Java side, as such, that includes methods bound using RegisterNatives. The option can be supplied multiple times. For example, -E JNI_OnLoad -E nativeMethod would exclude from the trace the JNI_OnLoad function call and any methods with the name nativeMethod.
-o path/output.json - is used to specify an output path where jnitrace will store all traced data. The information is stored in JSON format to allow later post-processing of the trace data.
-p path/to/script.js - the path provided is used to load a Frida script into the target process before the jnitrace script has loaded. This can be used for defeating anti-frida or anti-debugging code before jnitrace starts.
-a path/to/script.js - the path provided is used to load Frida script into the target process after jnitrace has been loaded.
--hide-data - used to reduce the quantity of output displayed in the console. This option will hide additional data that is displayed as hexdumps or as string de-references.
--ignore-env - using this option will hide all calls the app is making using the JNIEnv struct.
--ignore-vm - using this option will hide all calls the app is making using the JavaVM struct.
--aux <name=(string|bool|int)value> - used to pass custom parameters when spawning an application. For example --aux='uid=(int)10' will spawn the application for user 10 instead of default user 0.

Note

Remember frida-server must be running before running jnitrace. If the default instructions for installing frida have been followed, the following command will start the server ready for jnitrace:

adb shell /data/local/tmp/frida-server

API:

The engine that powers jnitrace is available as a separate project. That project allows you to import jnitrace to track individual JNI API calls, in a method familiar to using the Frida Interceptor to attach to functions and addresses.

import { JNIInterceptor } from "jnitrace-engine";

JNIInterceptor.attach("FindClass", {
    onEnter(args) {
        console.log("FindClass method called");
        this.className = Memory.readCString(args[1]);
    },
    onLeave(retval) {
        console.log("\tLoading Class:", this.className);
        console.log("\tClass ID:", retval.get());
    }
});

More information: https://github.com/chame1eon/jnitrace-engine

Building:

Building jnitrace from source requires that node first be installed. After installing node, the following commands need to be run:

npm install
npm run watch

npm run watch will run frida-compile in the background compiling the source to the output file, build/jnitrace.js. jnitrace.py loads from build/jnitrace.js by default, so no other changes are required to run the updates.

Output:

Like frida-trace, output is colored based on the API call thread.

Immediately below the thread ID in the display is the JNI API method name. Method names match exactly with those seen in the jni.h header file.

Subsequent lines contain a list of arguments indicated by a |-. After the |- characters are the argument type followed by the argument value. For jmethods, jfields and jclasses the Java type will be displayed in curly braces. This is dependent on jnitrace having seen the original method, field, or class lookup. For any methods passing buffers, jnitrace will extract the buffers from the arguments and display it as a hexdump below the argument value.

Return values are displayed at the bottom of the list as |= and will not be present for void methods.

If the backtrace is enabled, a Frida backtrace will be displayed below the method call. Please be aware, as per the Frida docs, the fuzzy backtrace is not always accurate and the accurate backtrace may provide limited results.

Details:

The goal of this project was to create a tool that could trace JNI API calls efficiently for most Android applications.

Unfortunately, the simplest approach of attaching to all function pointers in the JNIEnv structure overloads the application. It causes a crash based on the sheer number of function calls made by other unrelated libraries also using the same functions in libart.so.

To deal with that performance barrier, jnitrace creates a shadow JNIEnv that it can supply to libraries it wants to track. That JNIEnv contains a series of function trampolines that bounce the JNI API calls through some custom Frida NativeCallbacks to track the input and output of those functions.

The generic Frida API does a great job of providing a platform to build those function trampolines with minimal effort. However, that simple approach does not work for all of the JNIEnv API. The key problem with tracing all of the methods is the use of variadic arguments in the API. It is not possible to create the NativeCallback for these functions ahead of time, as it is not known beforehand all the different combinations of Java methods that will be called.

The solution is to monitor the process for calls to GetMethodID or GetStaticMethodID, used to look up method identifiers from the runtime. Once jnitrace sees a jmethodID lookup it has a known mapping of ID to method signature. Later, when a JNI Java method call is made, an initial NativeCallback is used to extract the method ID in the call. That method signature is then parsed to extract the method arguments. Once jnitrace has extracted the arguments in the method, it can dynamically create a NativeCallback for that method. That new NativeCallback is returned and a little bit of architecture specific shellcode deals with setting up the stack and registers to allow that call to run successfully. Those NativeCallbacks for specific methods are cached to allow the callback to run more efficiently if a method if called multiple times.

The other place where a simple NativeCallback is not sufficient for extracting the arguments from a method call, is for calls using a va_args pointer as the final argument. In this case jnitrace uses some code to extract the arguments from the pointer provided. Again this is architecture specific.

All data traced in these function calls is sent to the python console application that formats and displays it to the user.

Recommendations:

Most testing of this tool has been done on an Android x86_64 emulator running Marshmallow. Any issues experienced running on another device, please file an issue, but also, if possible, it is recommended to try running on a similar emulator.

Issues:

For any issues experienced running jnitrace please create an issue on GitHub. Please include the following information in the filed issue:

Device you were running on
Version of Frida you were using
Application you were running against
Any displayed error messages

TTDemo/jnitrace