JNI tips

JNI is the Java Native Interface. It defines a way for the bytecode that Android compiles from managed code (written in the Java or Kotlin programming languages) to interact with native code (written in C/C++). JNI is vendor-neutral, has support for loading code from dynamic shared libraries, and while cumbersome at times is reasonably efficient.

Note: Because Android compiles Kotlin to ART-friendly bytecode in a similar manner as the Java programming language, you can apply the guidance on this page to both the Kotlin and Java programming languages in terms of JNI architecture and its associated costs. To learn more, see Kotlin and Android.

If you're not already familiar with it, read through the Java Native Interface Specification to get a sense for how JNI works and what features are available. Some aspects of the interface aren't immediately obvious on first reading, so you may find the next few sections handy.

To browse global JNI references and see where global JNI references are created and deleted, use the JNI heap view in the Memory Profiler in Android Studio 3.2 and higher.

General tips

Try to minimize the footprint of your JNI layer. There are several dimensions to consider here. Your JNI solution should try to follow these guidelines (listed below by order of importance, beginning with the most important):

• Minimize marshalling of resources across the JNI layer. Marshalling across the JNI layer has non-trivial costs. Try to design an interface that minimizes the amount of data you need to marshall and the frequency with which you must marshall data.
• Avoid asynchronous communication between code written in a managed programming language and code written in C++ when possible. This will keep your JNI interface easier to maintain. You can typically simplify asynchronous UI updates by keeping the async update in the same language as the UI. For example, instead of invoking a C++ function from the UI thread in the Java code via JNI, it's better to do a callback between two threads in the Java programming language, with one of them making a blocking C++ call and then notifying the UI thread when the blocking call is complete.
• Minimize the number of threads that need to touch or be touched by JNI. If you do need to utilize thread pools in both the Java and C++ languages, try to keep JNI communication between the pool owners rather than between individual worker threads.
• Keep your interface code in a low number of easily identified C++ and Java source locations to facilitate future refactors. Consider using a JNI auto-generation library as appropriate.

JavaVM and JNIEnv

JNI defines two key data structures, "JavaVM" and "JNIEnv". Both of these are essentially pointers to pointers to function tables. (In the C++ version, they're classes with a pointer to a function table and a member function for each JNI function that indirects through the table.) The JavaVM provides the "invocation interface" functions, which allow you to create and destroy a JavaVM. In theory you can have multiple JavaVMs per process, but Android only allows one.

The JNIEnv provides most of the JNI functions. Your native functions all receive a JNIEnv as the first argument, except for @CriticalNative methods, see faster native calls.

The JNIEnv is used for thread-local storage. For this reason, you cannot share a JNIEnv between threads. If a piece of code has no other way to get its JNIEnv, you should share the JavaVM, and use GetEnv to discover the thread's JNIEnv. (Assuming it has one; see AttachCurrentThread below.)

The C declarations of JNIEnv and JavaVM are different from the C++ declarations. The "jni.h" include file provides different typedefs depending on whether it's included into C or C++. For this reason it's a bad idea to include JNIEnv arguments in header files included by both languages. (Put another way: if your header file requires #ifdef __cplusplus, you may have to do some extra work if anything in that header refers to JNIEnv.)

Threads

All threads are Linux threads, scheduled by the kernel. They're usually started from managed code (using Thread.start()), but they can also be created elsewhere and then attached to the JavaVM. For example, a thread started with pthread_create() or std::thread can be attached using the AttachCurrentThread() or AttachCurrentThreadAsDaemon() functions. Until a thread is attached, it has no JNIEnv, and cannot make JNI calls.

It's usually best to use Thread.start() to create any thread that needs to call in to Java code. Doing so will ensure that you have sufficient stack space, that you're in the correct ThreadGroup, and that you're using the same ClassLoader as your Java code. It's also easier to set the thread's name for debugging in Java than from native code (see pthread_setname_np() if you have a pthread_t or thread_t, and std::thread::native_handle() if you have a std::thread and want a pthread_t).

Attaching a natively-created thread causes a java.lang.Thread object to be constructed and added to the "main" ThreadGroup, making it visible to the debugger. Calling AttachCurrentThread() on an already-attached thread is a no-op.

Android does not suspend threads executing native code. If garbage collection is in progress, or the debugger has issued a suspend request, Android will pause the thread the next time it makes a JNI call.

Threads attached through JNI must call DetachCurrentThread() before they exit. If coding this directly is awkward, in Android 2.0 (Eclair) and higher you can use pthread_key_create() to define a destructor function that will be called before the thread exits, and call DetachCurrentThread() from there. (Use that key with pthread_setspecific() to store the JNIEnv in thread-local-storage; that way it'll be passed into your destructor as the argument.)

jclass, jmethodID, and jfieldID

If you want to access an object's field from native code, you would do the following:

• Get the class object reference for the class with FindClass
• Get the field ID for the field with GetFieldID
• Get the contents of the field with something appropriate, such as GetIntField

Similarly, to call a method, you'd first get a class object reference and then a method ID. The IDs are often just pointers to internal runtime data structures. Looking them up may require several string comparisons, but once you have them the actual call to get the field or invoke the method is very quick.

If performance is important, it's useful to look the values up once and cache the results in your native code. Because there is a limit of one JavaVM per process, it's reasonable to store this data in a static local structure.

The class references, field IDs, and method IDs are guaranteed valid until the class is unloaded. Classes are only unloaded if all classes associated with a ClassLoader can be garbage collected, which is rare but will not be impossible in Android. Note however that the jclass is a class reference and must be protected with a call to NewGlobalRef (see the next section).

If you would like to cache the IDs when a class is loaded, and automatically re-cache them if the class is ever unloaded and reloaded, the correct way to initialize the IDs is to add a piece of code that looks like this to the appropriate class:

companion object {
    /*
     * We use a static class initializer to allow the native code to cache some
     * field offsets. This native function looks up and caches interesting
     * class/field/method IDs. Throws on failure.
     */
    private external fun nativeInit()

    init {
        nativeInit()
    }
}

    /*
     * We use a class initializer to allow the native code to cache some
     * field offsets. This native function looks up and caches interesting
     * class/field/method IDs. Throws on failure.
     */
    private static native void nativeInit();

    static {
        nativeInit();
    }

Create a nativeClassInit method in your C/C++ code that performs the ID lookups. The code will be executed once, when the class is initialized. If the class is ever unloaded and then reloaded, it will be executed again.

Local and global references

Every argument passed to a native method, and almost every object returned by a JNI function is a "local reference". This means that it's valid for the duration of the current native method in the current thread. Even if the object itself continues to live on after the native method returns, the reference is not valid.

This applies to all sub-classes of jobject, including jclass, jstring, and jarray. (The runtime will warn you about most reference mis-uses when extended JNI checks are enabled.)

The only way to get non-local references is via the functions NewGlobalRef and NewWeakGlobalRef.

If you want to hold on to a reference for a longer period, you must use a "global" reference. The NewGlobalRef function takes the local reference as an argument and returns a global one. The global reference is guaranteed to be valid until you call DeleteGlobalRef.

This pattern is commonly used when caching a jclass returned from FindClass, e.g.:

jclass localClass = env->FindClass("MyClass");
jclass globalClass = reinterpret_cast<jclass>(env->NewGlobalRef(localClass));

All JNI methods accept both local and global references as arguments. It's possible for references to the same object to have different values. For example, the return values from consecutive calls to NewGlobalRef on the same object may be different. To see if two references refer to the same object, you must use the IsSameObject function. Never compare references with == in native code.

One consequence of this is that you must not assume object references are constant or unique in native code. The value representing an object may be different from one invocation of a method to the next, and it's possible that two different objects could have the same value on consecutive calls. Do not use jobject values as keys.

Programmers are required to "not excessively allocate" local references. In practical terms this means that if you're creating large numbers of local references, perhaps while running through an array of objects, you should free them manually with DeleteLocalRef instead of letting JNI do it for you. The implementation is only required to reserve slots for 16 local references, so if you need more than that you should either delete as you go or use EnsureLocalCapacity/PushLocalFrame to reserve more.

Note that jfieldIDs and jmethodIDs are opaque types, not object references, and should not be passed to NewGlobalRef. The raw data pointers returned by functions like GetStringUTFChars and GetByteArrayElements are also not objects. (They may be passed between threads, and are valid until the matching Release call.)

One unusual case deserves separate mention. If you attach a native thread with AttachCurrentThread, the code you are running will never automatically free local references until the thread detaches. Any local references you create will have to be deleted manually. In general, any native code that creates local references in a loop probably needs to do some manual deletion.

Be careful using global references. Global references can be unavoidable, but they are difficult to debug and can cause difficult-to-diagnose memory (mis)behaviors. All else being equal, a solution with fewer global references is probably better.

UTF-8 and UTF-16 strings

The Java programming language uses UTF-16. For convenience, JNI provides methods that work with Modified UTF-8 as well. The modified encoding is useful for C code because it encodes \u0000 as 0xc0 0x80 instead of 0x00. The nice thing about this is that you can count on having C-style zero-terminated strings, suitable for use with standard libc string functions. The down side is that you cannot pass arbitrary UTF-8 data to JNI and expect it to work correctly.

To get the UTF-16 representation of a String, use GetStringChars. Note that UTF-16 strings are not zero-terminated, and \u0000 is allowed, so you need to hang on to the string length as well as the jchar pointer.

Don't forget to Release the strings you Get. The string functions return jchar* or jbyte*, which are C-style pointers to primitive data rather than local references. They are guaranteed valid until Release is called, which means they are not released when the native method returns.

Data passed to NewStringUTF must be in Modified UTF-8 format. A common mistake is reading character data from a file or network stream and handing it to NewStringUTF without filtering it. Unless you know the data is valid MUTF-8 (or 7-bit ASCII, which is a compatible subset), you need to strip out invalid characters or convert them to proper Modified UTF-8 form. If you don't, the UTF-16 conversion is likely to provide unexpected results. CheckJNI—which is on by default for emulators—scans strings and aborts the VM if it receives invalid input.

Before Android 8, it was usually faster to operate with UTF-16 strings as Android did not require a copy in GetStringChars, whereas GetStringUTFChars required an allocation and a conversion to UTF-8. Android 8 changed the String representation to use 8 bits per character for ASCII strings (to save memory) and started to use a moving garbage collector. These features greatly reduce the number of cases where ART can provide a pointer to the String data without making a copy, even for GetStringCritical. However, if most strings processed by the code are short, it is possible to avoid the allocation and deallocation in most cases by using a stack-allocated buffer and GetStringRegion or GetStringUTFRegion. For example:

    constexpr size_t kStackBufferSize = 64u;
    jchar stack_buffer[kStackBufferSize];
    std::unique_ptr<jchar[]> heap_buffer;
    jchar* buffer = stack_buffer;
    jsize length = env->GetStringLength(str);
    if (length > kStackBufferSize) {
      heap_buffer.reset(new jchar[length]);
      buffer = heap_buffer.get();
    }
    env->GetStringRegion(str, 0, length, buffer);
    process_data(buffer, length);

Primitive arrays

JNI provides functions for accessing the contents of array objects. While arrays of objects must be accessed one entry at a time, arrays of primitives can be read and written directly as if they were declared in C.

To make the interface as efficient as possible without constraining the VM implementation, the Get<PrimitiveType>ArrayElements family of calls allows the runtime to either return a pointer to the actual elements, or allocate some memory and make a copy. Either way, the raw pointer returned is guaranteed to be valid until the corresponding Release call is issued (which implies that, if the data wasn't copied, the array object will be pinned down and can't be relocated as part of compacting the heap). You must Release every array you Get. Also, if the Get call fails, you must ensure that your code doesn't try to Release a NULL pointer later.

You can determine whether or not the data was copied by passing in a non-NULL pointer for the isCopy argument. This is rarely useful.

The Release call takes a mode argument that can have one of three values. The actions performed by the runtime depend upon whether it returned a pointer to the actual data or a copy of it:

• 0
  • Actual: the array object is un-pinned.
  • Copy: data is copied back. The buffer with the copy is freed.
• JNI_COMMIT
  • Actual: does nothing.
  • Copy: data is copied back. The buffer with the copy is not freed.
• JNI_ABORT
  • Actual: the array object is un-pinned. Earlier writes are not aborted.
  • Copy: the buffer with the copy is freed; any changes to it are lost.

One reason for checking the isCopy flag is to know if you need to call Release with JNI_COMMIT after making changes to an array — if you're alternating between making changes and executing code that uses the contents of the array, you may be able to skip the no-op commit. Another possible reason for checking the flag is for efficient handling of JNI_ABORT. For example, you might want to get an array, modify it in place, pass pieces to other functions, and then discard the changes. If you know that JNI is making a new copy for you, there's no need to create another "editable" copy. If JNI is passing you the original, then you do need to make your own copy.

It is a common mistake (repeated in example code) to assume that you can skip the Release call if *isCopy is false. This is not the case. If no copy buffer was allocated, then the original memory must be pinned down and can't be moved by the garbage collector.

Also note that the JNI_COMMIT flag does not release the array, and you will need to call Release again with a different flag eventually.

Region calls

There is an alternative to calls like Get<Type>ArrayElements and GetStringChars that may be very helpful when all you want to do is copy data in or out. Consider the following:

    jbyte* data = env->GetByteArrayElements(array, NULL);
    if (data != NULL) {
        memcpy(buffer, data, len);
        env->ReleaseByteArrayElements(array, data, JNI_ABORT);
    }

This grabs the array, copies the first len byte elements out of it, and then releases the array. Depending upon the implementation, the Get call will either pin or copy the array contents. The code copies the data (for perhaps a second time), then calls Release; in this case JNI_ABORT ensures there's no chance of a third copy.

One can accomplish the same thing more simply:

    env->GetByteArrayRegion(array, 0, len, buffer);

This has several advantages:

• Requires one JNI call instead of 2, reducing overhead.
• Doesn't require pinning or extra data copies.
• Reduces the risk of programmer error — no risk of forgetting to call Release after something fails.

Similarly, you can use the Set<Type>ArrayRegion call to copy data into an array, and GetStringRegion or GetStringUTFRegion to copy characters out of a String.

Exceptions

You must not call most JNI functions while an exception is pending. Your code is expected to notice the exception (via the function's return value, ExceptionCheck, or ExceptionOccurred) and return, or clear the exception and handle it.

The only JNI functions that you are allowed to call while an exception is pending are:

• DeleteGlobalRef
• DeleteLocalRef
• DeleteWeakGlobalRef
• ExceptionCheck
• ExceptionClear
• ExceptionDescribe
• ExceptionOccurred
• MonitorExit
• PopLocalFrame
• PushLocalFrame
• Release<PrimitiveType>ArrayElements
• ReleasePrimitiveArrayCritical
• ReleaseStringChars
• ReleaseStringCritical
• ReleaseStringUTFChars

Many JNI calls can throw an exception, but often provide a simpler way of checking for failure. For example, if NewString returns a non-NULL value, you don't need to check for an exception. However, if you call a method (using a function like CallObjectMethod), you must always check for an exception, because the return value is not going to be valid if an exception was thrown.

Note that exceptions thrown by managed code do not unwind native stack frames. (And C++ exceptions, generally discouraged on Android, must not be thrown across the JNI transition boundary from C++ code to managed code.) The JNI Throw and ThrowNew instructions just set an exception pointer in the current thread. Upon returning to managed from native code, the exception will be noted and handled appropriately.

Native code can "catch" an exception by calling ExceptionCheck or ExceptionOccurred, and clear it with ExceptionClear. As usual, discarding exceptions without handling them can lead to problems.

There are no built-in functions for manipulating the Throwable object itself, so if you want to (say) get the exception string you will need to find the Throwable class, look up the method ID for getMessage "()Ljava/lang/String;", invoke it, and if the result is non-NULL use GetStringUTFChars to get something you can hand to printf(3) or equivalent.

Extended checking

JNI does very little error checking. Errors usually result in a crash. Android also offers a mode called CheckJNI, where the JavaVM and JNIEnv function table pointers are switched to tables of functions that perform an extended series of checks before calling the standard implementation.

The additional checks include:

• Arrays: attempting to allocate a negative-sized array.
• Bad pointers: passing a bad jarray/jclass/jobject/jstring to a JNI call, or passing a NULL pointer to a JNI call with a non-nullable argument.
• Class names: passing anything but the “java/lang/String” style of class name to a JNI call.
• Critical calls: making a JNI call between a “critical” get and its corresponding release.
• Direct ByteBuffers: passing bad arguments to NewDirectByteBuffer.
• Exceptions: making a JNI call while there’s an exception pending.
• JNIEnv*s: using a JNIEnv* from the wrong thread.
• jfieldIDs: using a NULL jfieldID, or using a jfieldID to set a field to a value of the wrong type (trying to assign a StringBuilder to a String field, say), or using a jfieldID for a static field to set an instance field or vice versa, or using a jfieldID from one class with instances of another class.
• jmethodIDs: using the wrong kind of jmethodID when making a Call*Method JNI call: incorrect return type, static/non-static mismatch, wrong type for ‘this’ (for non-static calls) or wrong class (for static calls).
• References: using DeleteGlobalRef/DeleteLocalRef on the wrong kind of reference.
• Release modes: passing a bad release mode to a release call (something other than 0, JNI_ABORT, or JNI_COMMIT).
• Type safety: returning an incompatible type from your native method (returning a StringBuilder from a method declared to return a String, say).
• UTF-8: passing an invalid Modified UTF-8 byte sequence to a JNI call.

(Accessibility of methods and fields is still not checked: access restrictions don't apply to native code.)

There are several ways to enable CheckJNI.

If you’re using the emulator, CheckJNI is on by default.

If you have a rooted device, you can use the following sequence of commands to restart the runtime with CheckJNI enabled:

adb shell stop
adb shell setprop dalvik.vm.checkjni true
adb shell start

In either of these cases, you’ll see something like this in your logcat output when the runtime starts:

D AndroidRuntime: CheckJNI is ON

If you have a regular device, you can use the following command:

adb shell setprop debug.checkjni 1

This won’t affect already-running apps, but any app launched from that point on will have CheckJNI enabled. (Change the property to any other value or simply rebooting will disable CheckJNI again.) In this case, you’ll see something like this in your logcat output the next time an app starts:

D Late-enabling CheckJNI

You can also set the android:debuggable attribute in your application's manifest to turn on CheckJNI just for your app. Note that the Android build tools will do this automatically for certain build types.

Native libraries

You can load native code from shared libraries with the standard System.loadLibrary.

In practice, older versions of Android had bugs in PackageManager that caused installation and update of native libraries to be unreliable. The ReLinker project offers workarounds for this and other native library loading problems.

Call System.loadLibrary (or ReLinker.loadLibrary) from a static class initializer. The argument is the "undecorated" library name, so to load libfubar.so you would pass in "fubar".

If you have only one class with native methods, it makes sense for the call to System.loadLibrary to be in a static initializer for that class. Otherwise you might want to make the call from Application so you know that the library is always loaded, and always loaded early.

There are two ways that the runtime can find your native methods. You can either explicitly register them with RegisterNatives, or you can let the runtime look them up dynamically with dlsym. The advantages of RegisterNatives are that you get up-front checking that the symbols exist, plus you can have smaller and faster shared libraries by not exporting anything but JNI_OnLoad. The advantage of letting the runtime discover your functions is that it's slightly less code to write.

To use RegisterNatives:

• Provide a JNIEXPORT jint JNI_OnLoad(JavaVM* vm, void* reserved) function.
• In your JNI_OnLoad, register all of your native methods using RegisterNatives.
• Build with a version script (preferred) or use -fvisibility=hidden so that only your JNI_OnLoad is exported from your library. This produces faster and smaller code, and avoids potential collisions with other libraries loaded into your app (but it creates less useful stack traces if your app crashes in native code).

The static initializer should look like this:

companion object {
    init {
        System.loadLibrary("fubar")
    }
}

static {
    System.loadLibrary("fubar");
}

The JNI_OnLoad function should look something like this if written in C++:

JNIEXPORT jint JNI_OnLoad(JavaVM* vm, void* reserved) {
    JNIEnv* env;
    if (vm->GetEnv(reinterpret_cast<void**>(&env), JNI_VERSION_1_6) != JNI_OK) {
        return JNI_ERR;
    }

    // Find your class. JNI_OnLoad is called from the correct class loader context for this to work.
    jclass c = env->FindClass("com/example/app/package/MyClass");
    if (c == nullptr) return JNI_ERR;

    // Register your class' native methods.
    static const JNINativeMethod methods[] = {
        {"nativeFoo", "()V", reinterpret_cast<void*>(nativeFoo)},
        {"nativeBar", "(Ljava/lang/String;I)Z", reinterpret_cast<void*>(nativeBar)},
    };
    int rc = env->RegisterNatives(c, methods, sizeof(methods)/sizeof(JNINativeMethod));
    if (rc != JNI_OK) return rc;

    return JNI_VERSION_1_6;
}

To instead use "discovery" of native methods, you need to name them in a specific way (see the JNI spec for details). This means that if a method signature is wrong, you won't know about it until the first time the method is actually invoked.

Any FindClass calls made from JNI_OnLoad will resolve classes in the context of the class loader that was used to load the shared library. When called from other contexts, FindClass uses the class loader associated with the method at the top of the Java stack, or if there isn't one (because the call is from a native thread that was just attached) it uses the "system" class loader. The system class loader does not know about your application's classes, so you won't be able to look up your own classes with FindClass in that context. This makes JNI_OnLoad a convenient place to look up and cache classes: once you have a valid jclass global reference you can use it from any attached thread.

Faster native calls with @FastNative and @CriticalNative

Native methods can be annotated with @FastNative or @CriticalNative (but not both) to speed up transitions between managed and native code. However, these annotations come with certain changes in behavior that need to be carefully considered before use. While we briefly mention these changes below, please refer to the documentation for the details.

The @CriticalNative annotation can be applied only to native methods that do not use managed objects (in parameters or return values, or as an implicit this), and this annotation changes the JNI transition ABI. The native implementation must exclude the JNIEnv and jclass parameters from its function signature.

While executing a @FastNative or @CriticalNative method, the garbage collection cannot suspend the thread for essential work and may become blocked. Do not use these annotations for long-running methods, including usually-fast, but generally unbounded, methods. In particular, the code should not perform significant I/O operations or acquire native locks that can be held for a long time.

These annotations were implemented for system use since Android 8 and became CTS-tested public API in Android 14. These optimizations are likely to work also on Android 8-13 devices (albeit without the strong CTS guarantees) but the dynamic lookup of native methods is supported only on Android 12+, the explicit registration with JNI RegisterNatives is strictly required for running on Android versions 8-11. These annotations are ignored on Android 7-, the ABI mismatch for @CriticalNative would lead to wrong argument marshalling and likely crashes.

For performance critical methods that need these annotations, it is strongly recommended to explicitly register the method(s) with JNI RegisterNatives instead of relying on the name-based "discovery" of native methods. To get optimal app startup performance, it is recommended to include callers of @FastNative or @CriticalNative methods in the baseline profile. Since Android 12, a call to a @CriticalNative native method from a compiled managed method is almost as cheap as a non-inline call in C/C++ as long as all arguments fit into registers (for example up to 8 integral and up to 8 floating point arguments on arm64).

Sometimes it can be preferable to split a native method into two, a very fast method that can fail and another one that handles the slow cases. For example:

fun writeInt(nativeHandle: Long, value: Int) {
    // A fast buffered write with a `@CriticalNative` method should succeed most of the time.
    if (!nativeTryBufferedWriteInt(nativeHandle, value)) {
        // If the buffered write failed, we need to use the slow path that can perform
        // significant I/O and can even throw an `IOException`.
        nativeWriteInt(nativeHandle, value)
    }
}

@CriticalNative
external fun nativeTryBufferedWriteInt(nativeHandle: Long, value: Int): Boolean

external fun nativeWriteInt(nativeHandle: Long, value: Int)

void writeInt(long nativeHandle, int value) {
    // A fast buffered write with a `@CriticalNative` method should succeed most of the time.
    if (!nativeTryBufferedWriteInt(nativeHandle, value)) {
        // If the buffered write failed, we need to use the slow path that can perform
        // significant I/O and can even throw an `IOException`.
        nativeWriteInt(nativeHandle, value);
    }
}

@CriticalNative
static native boolean nativeTryBufferedWriteInt(long nativeHandle, int value);

static native void nativeWriteInt(long nativeHandle, int value);

64-bit considerations

To support architectures that use 64-bit pointers, use a long field rather than an int when storing a pointer to a native structure in a Java field.

Unsupported features/backwards compatibility

All JNI 1.6 features are supported, with the following exception:

• DefineClass is not implemented. Android does not use Java bytecodes or class files, so passing in binary class data doesn't work.

For backward compatibility with older Android releases, you may need to be aware of:

• Dynamic lookup of native functions

  Until Android 2.0 (Eclair), the '$' character was not properly converted to "_00024" during searches for method names. Working around this requires using explicit registration or moving the native methods out of inner classes.

• Detaching threads

  Until Android 2.0 (Eclair), it was not possible to use a pthread_key_create destructor function to avoid the "thread must be detached before exit" check. (The runtime also uses a pthread key destructor function, so it'd be a race to see which gets called first.)

• Weak global references

  Until Android 2.2 (Froyo), weak global references were not implemented. Older versions will vigorously reject attempts to use them. You can use the Android platform version constants to test for support.

  Until Android 4.0 (Ice Cream Sandwich), weak global references could only be passed to NewLocalRef, NewGlobalRef, and DeleteWeakGlobalRef. (The spec strongly encourages programmers to create hard references to weak globals before doing anything with them, so this should not be at all limiting.)

  From Android 4.0 (Ice Cream Sandwich) on, weak global references can be used like any other JNI references.

• Local references

  Until Android 4.0 (Ice Cream Sandwich), local references were actually direct pointers. Ice Cream Sandwich added the indirection necessary to support better garbage collectors, but this means that lots of JNI bugs are undetectable on older releases. See JNI Local Reference Changes in ICS for more details.

  In Android versions prior to Android 8.0, the number of local references is capped at a version-specific limit. Beginning in Android 8.0, Android supports unlimited local references.

• Determining reference type with GetObjectRefType

  Until Android 4.0 (Ice Cream Sandwich), as a consequence of the use of direct pointers (see above), it was impossible to implement GetObjectRefType correctly. Instead we used a heuristic that looked through the weak globals table, the arguments, the locals table, and the globals table in that order. The first time it found your direct pointer, it would report that your reference was of the type it happened to be examining. This meant, for example, that if you called GetObjectRefType on a global jclass that happened to be the same as the jclass passed as an implicit argument to your static native method, you'd get JNILocalRefType rather than JNIGlobalRefType.

• @FastNative and @CriticalNative

  Up to Android 7, these optimization annotations were ignored. The ABI mismatch for @CriticalNative would lead to wrong argument marshalling and likely crashes.

  The dynamic lookup of native functions for @FastNative and @CriticalNative methods was unimplemented in Android 8-10 and contains known bugs in Android 11. Using these optimizations without explicit registration with JNI RegisterNatives is likely to lead to crashes on Android 8-11.

• FindClass throws ClassNotFoundException

  For backward compatibility, Android throws ClassNotFoundException instead of NoClassDefFoundError when a class isn't found by FindClass. This behavior is consistent with the Java reflection API Class.forName(name).

FAQ: Why do I get UnsatisfiedLinkError?

When working on native code it's not uncommon to see a failure like this:

java.lang.UnsatisfiedLinkError: Library foo not found

In some cases it means what it says — the library wasn't found. In other cases the library exists but couldn't be opened by dlopen(3), and the details of the failure can be found in the exception's detail message.

Common reasons why you might encounter "library not found" exceptions:

• The library doesn't exist or isn't accessible to the app. Use adb shell ls -l <path> to check its presence and permissions.
• The library wasn't built with the NDK. This can result in dependencies on functions or libraries that don't exist on the device.

Another class of UnsatisfiedLinkError failures looks like:

java.lang.UnsatisfiedLinkError: myfunc
        at Foo.myfunc(Native Method)
        at Foo.main(Foo.java:10)

In logcat, you'll see:

W/dalvikvm(  880): No implementation found for native LFoo;.myfunc ()V

This means that the runtime tried to find a matching method but was unsuccessful. Some common reasons for this are:

• The library isn't getting loaded. Check the logcat output for messages about library loading.
• The method isn't being found due to a name or signature mismatch. This is commonly caused by:
  • For lazy method lookup, failing to declare C++ functions with extern "C" and appropriate visibility (JNIEXPORT). Note that prior to Ice Cream Sandwich, the JNIEXPORT macro was incorrect, so using a new GCC with an old jni.h won't work. You can use arm-eabi-nm to see the symbols as they appear in the library; if they look mangled (something like _Z15Java_Foo_myfuncP7_JNIEnvP7_jclass rather than Java_Foo_myfunc), or if the symbol type is a lowercase 't' rather than an uppercase 'T', then you need to adjust the declaration.
  • For explicit registration, minor errors when entering the method signature. Make sure that what you're passing to the registration call matches the signature in the log file. Remember that 'B' is byte and 'Z' is boolean. Class name components in signatures start with 'L', end with ';', use '/' to separate package/class names, and use '$' to separate inner-class names (Ljava/util/Map$Entry;, say).

Using javah to automatically generate JNI headers may help avoid some problems.

FAQ: Why didn't FindClass find my class?

(Most of this advice applies equally well to failures to find methods with GetMethodID or GetStaticMethodID, or fields with GetFieldID or GetStaticFieldID.)

Make sure that the class name string has the correct format. JNI class names start with the package name and are separated with slashes, such as java/lang/String. If you're looking up an array class, you need to start with the appropriate number of square brackets and must also wrap the class with 'L' and ';', so a one-dimensional array of String would be [Ljava/lang/String;. If you're looking up an inner class, use '$' rather than '.'. In general, using javap on the .class file is a good way to find out the internal name of your class.

If you enable code shrinking, make sure that you configure which code to keep. Configuring proper keep rules is important because the code shrinker might otherwise remove classes, methods, or fields that are only used from JNI.

If the class name looks right, you could be running into a class loader issue. FindClass wants to start the class search in the class loader associated with your code. It examines the call stack, which will look something like:

    Foo.myfunc(Native Method)
    Foo.main(Foo.java:10)

The topmost method is Foo.myfunc. FindClass finds the ClassLoader object associated with the Foo class and uses that.

This usually does what you want. You can get into trouble if you create a thread yourself (perhaps by calling pthread_create and then attaching it with AttachCurrentThread). Now there are no stack frames from your application. If you call FindClass from this thread, the JavaVM will start in the "system" class loader instead of the one associated with your application, so attempts to find app-specific classes will fail.

There are a few ways to work around this:

• Do your FindClass lookups once, in JNI_OnLoad, and cache the class references for later use. Any FindClass calls made as part of executing JNI_OnLoad will use the class loader associated with the function that called System.loadLibrary (this is a special rule, provided to make library initialization more convenient). If your app code is loading the library, FindClass will use the correct class loader.
• Pass an instance of the class into the functions that need it, by declaring your native method to take a Class argument and then passing Foo.class in.
• Cache a reference to the ClassLoader object somewhere handy, and issue loadClass calls directly. This requires some effort.

FAQ: How do I share raw data with native code?

You may find yourself in a situation where you need to access a large buffer of raw data from both managed and native code. Common examples include manipulation of bitmaps or sound samples. There are two basic approaches.

You can store the data in a byte[]. This allows very fast access from managed code. On the native side, however, you're not guaranteed to be able to access the data without having to copy it. In some implementations, GetByteArrayElements and GetPrimitiveArrayCritical will return actual pointers to the raw data in the managed heap, but in others it will allocate a buffer on the native heap and copy the data over.

The alternative is to store the data in a direct byte buffer. These can be created with java.nio.ByteBuffer.allocateDirect, or the JNI NewDirectByteBuffer function. Unlike regular byte buffers, the storage is not allocated on the managed heap, and can always be accessed directly from native code (get the address with GetDirectBufferAddress). Depending on how direct byte buffer access is implemented, accessing the data from managed code can be very slow.

The choice of which to use depends on two factors:

1. Will most of the data accesses happen from code written in Java or in C/C++?
2. If the data is eventually being passed to a system API, what form must it be in? (For example, if the data is eventually passed to a function that takes a byte[], doing processing in a direct ByteBuffer might be unwise.)

If there's no clear winner, use a direct byte buffer. Support for them is built directly into JNI, and performance should improve in future releases.