Skip to content

Latest commit

 

History

History
306 lines (260 loc) · 11.9 KB

ImplementingNativeMethodsInJavaWithSVM.md

File metadata and controls

306 lines (260 loc) · 11.9 KB

Implementing Native Methods in Java with Substrate VM

Substrate VM can be used to implement low-level system operations in Java and make them available via JNI to Java code executing on a standard JVM. As a result one can use the same language to write the application logic as well as the system calls.

Note that this document describes the opposite of what is commonly done via JNI: Usually low-level system operations are implemented in C and invoked from Java using JNI. If you are interested in how Substrate VM supports the common use case, read the documentation about Substrate VM JNI support instead.

Create a Shared Library

First of all one has to use the native-image command to generate a shared library with some JNI-compatible entry points. Let's start with the Java code:

package org.pkg.implnative;

import org.graalvm.nativeimage.c.function.CEntryPoint;
import org.graalvm.word.Pointer;

public final class NativeImpl {
    @CEntryPoint(name = "Java_org_pkg_apinative_Native_add")
    public static int add(Pointer jniEnv, Pointer clazz, @CEntryPoint.IsolateThreadContext long isolateId, int a, int b) {
        return a + b;
    }
}

After being processed by the native-image command the code exposes a C function Java_org_pkg_apinative_Native_add (the name follows conventions of JNI that will be handy later) and a SubstrateVM signature typical for JNI methods. The first parameter is a reference to JNIEnv* value, the second parameter is a reference to the jclass value for the class declaring the method. The third parameter is a portable (e.g. long) identifier of the SubstrateVM isolatethread. The rest of the parameters are the actual parameters of the Java Native.add method described in the next section. Compile the code with shared option on:

$GRAALVM/bin/native-image --shared -H:Name=libnativeimpl -cp nativeimpl

and the libnativeimpl.so is generated. We are ready to use it from standard Java code.

Bind Java Native Method

Now we need another Java class to use the native library generated in the previous step:

package org.pkg.apinative;

public final class Native {
    private static native int add(long isolateThreadId, int a, int b);
}

the package name of the class as well as name of the method has to correspond (after the JNI mangling) to the name of the @CEntryPoint introduced previously. The first argument is a portable (e.g. long) identifier of the SubstrateVM isolate thread. The rest of the arguments matches the parameters of the entry point.

Loading the Native Library

The next step is to bind the JDK with the generated .so library - e.g. make sure the implementation of the native Native.add method is loaded. Simple load or loadLibrary calls will do:

public static void main(String[] args) {
    System.loadLibrary("nativeimpl");
    // ...
}

under the assumption your LD_LIBRARY_PATH environment variable is specified or java.library.path Java property is properly set.

Initializing the Substrate VM

Before making calls to the Native.add method, we need to create a Substrate VM isolate. Substrate VM provides special built-in to allow that: CEntryPoint.Builtin.CREATE_ISOLATE. Define another method along your other existing @CEntryPoint methods. Let it return IsolateThread and take no parameters:

public final class NativeImpl {
    @CEntryPoint(name = "Java_org_pkg_apinative_Native_createIsolate", builtin=CEntryPoint.Builtin.CREATE_ISOLATE)
    public static native IsolateThread createIsolate();
}

SubstrateVM then generates default native implementation of the method into the final .so library. The method initializes the Substrate VM runtime and returns a portable identification - e.g. long to hold an instance of a Substrate VM isolatethread. The isolate thread can then be used for multiple invocations of the native part of our code:

package org.pkg.apinative;

public final class Native {
    public static void main(String[] args) {
        System.loadLibrary("nativeimpl");

        long isolateThread = createIsolate();

        System.out.println("2 + 40 = " + add(isolateThread, 2, 40));
        System.out.println("12 + 30 = " + add(isolateThread, 12, 30));
        System.out.println("20 + 22 = " + add(isolateThread, 20, 22));
    }

    private static native int add(long isolateThread, int a, int b);
    private static native long createIsolate();
}

The standard JVM is started. It initializes a Substrate VM isolate, attaches current thread to the isolate and the universal answer 42 is then computed three times inside of the isolate.

Calling JVM from Native Java

There is a detailed tutorial on the C interface of Substrate VM. Rather than repeating it here, let's apply some of its principles to a related topic, the implementation of native methods. Let's make a callback to JVM!

In the classical setup, when C needs to call into JVM, it uses jni.h header file. The file defines essential JVM structures (like JNIEnv) as well as functions one can invoke to inspect classes, access fields, and call methods in the JVM. In order to call these functions from our NativeImpl class in our example, we need to define appropriate Java API wrappers of the jni.h concepts:

@CContext(JNIHeaderDirectives.class)
@CStruct(value = "JNIEnv_", addStructKeyword = true)
interface JNIEnvironment extends PointerBase {
    @CField("functions")
    JNINativeInterface getFunctions();
}

@CPointerTo(JNIEnvironment.class)
interface JNIEnvironmentPointer extends PointerBase {
    JNIEnvironment read();
    void write(JNIEnvironment value);
}

@CContext(JNIHeaderDirectives.class)
@CStruct(value = "JNINativeInterface_", addStructKeyword = true)
interface JNINativeInterface extends PointerBase {
    @CField
    GetMethodId getGetStaticMethodID();

    @CField
    CallStaticVoidMethod getCallStaticVoidMethodA();
}

interface GetMethodId extends CFunctionPointer {
    @InvokeCFunctionPointer
    JMethodID find(JNIEnvironment env, JClass clazz, CCharPointer name, CCharPointer sig);
}

interface JObject extends PointerBase {
}

interface CallStaticVoidMethod extends CFunctionPointer {
    @InvokeCFunctionPointer
    void call(JNIEnvironment env, JClass cls, JMethodID methodid, JValue args);
}

interface JClass extends PointerBase {
}
interface JMethodID extends PointerBase {
}

If we leave aside the meaning of JNIHeaderDirectives for now, the rest of the interfaces is a type-safe representation of the C pointers found in the jni.h file. JClass, JMethodID, JObject are all pointers. Thanks to the above definitions we now have Java interfaces to represent instances of these objects in our native Java code in a type safe way.

The core part of any JNI API is the set of functions one can call when talking to the JVM. There are dozens of them, but in the JNINativeInterface definition we just define wrappers for those few that we need in our example. Again, we give them proper types, so in our native Java code we can use GetMethodId.find(...), CallStaticVoidMethod.call(...), etc. In addition, there is another important part missing in the puzzle - the jvalue union type wrapping all the possible Java primitive and object types. Here comes definition of its getters and setters:

@CContext(JNIHeaderDirectives.class)
@CStruct("jvalue")
interface JValue extends PointerBase {
    @CField boolean z();
    @CField byte b();
    @CField char c();
    @CField short s();
    @CField int i();
    @CField long j();
    @CField float f();
    @CField double d();
    @CField JObject l();


    @CField void z(boolean b);
    @CField void b(byte b);
    @CField void c(char ch);
    @CField void s(short s);
    @CField void i(int i);
    @CField void j(long l);
    @CField void f(float f);
    @CField void d(double d);
    @CField void l(JObject obj);

    JValue addressOf(int index);
}

The addressOf method is a special Substrate VM construct used to perform C pointer arithmetics. Given a pointer one can treat it as initial element of an array, then for example, use addressOf(1) to access the subsequent element. With this we have all the API we need to make a callback: let's redefine the previously introduced NativeImpl.add method to accept properly typed pointers, and then use these pointers to invoke a JVM method before computing the sum of a + b:

@CEntryPoint(name = "Java_org_pkg_apinative_Native_add")
static int add(JNIEnvironment env, JClass clazz, @CEntryPoint.IsolateThreadContext long isolateThreadId, int a, int b) {
    JNINativeInterface fn = env.getFunctions();

    try (
        CTypeConversion.CCharPointerHolder name = CTypeConversion.toCString("hello");
        CTypeConversion.CCharPointerHolder sig = CTypeConversion.toCString("(ZCBSIJFD)V");
    ) {
        JMethodID helloId = fn.getGetStaticMethodID().find(env, clazz, name.get(), sig.get());

        JValue args = StackValue.get(8, JValue.class);
        args.addressOf(0).z(false);
        args.addressOf(1).c('A');
        args.addressOf(2).b((byte)22);
        args.addressOf(3).s((short)33);
        args.addressOf(4).i(39);
        args.addressOf(5).j(Long.MAX_VALUE / 2l);
        args.addressOf(6).f((float) Math.PI);
        args.addressOf(7).d(Math.PI);
        fn.getCallStaticVoidMethodA().call(env, clazz, helloId, args);
    }

    return a + b;
}

The above example seeks a static method hello and invokes it with eight JValue parameters in an array reserved by StackValue.get on the stack. Individual parameters are accessed by use of the addressOf operator and filled with appropriate primitive values before the call happens. The method hello is defined in the class Native and just prints values of all parameters to verify they are properly propagated from the NativeImpl.add caller:

public class Native {
    public static void hello(boolean z, char c, byte b, short s, int i, long j, float f, double d) {
        System.err.println("Hi, I have just been called back!");
        System.err.print("With: " + z + " " + c + " " + b + " " + s);
        System.err.println(" and: " + i + " " + j + " " + f + " " + d);
    }

There is just one final piece to explain: the JNIHeaderDirectives. Substrate VM C interface needs to understand the layout of the C structures. It needs to know at which offset of JNINativeInterface structure it can find the pointer to GetMethodId function. To do so, it needs jni.h and additional files during compilation. One can specify them by @CContext annotation and implementation of its Directives:

final class JNIHeaderDirectives implements CContext.Directives {
    @Override
    public List<String> getOptions() {
        File[] jnis = findJNIHeaders();
        return Arrays.asList("-I" + jnis[0].getParent(), "-I" + jnis[1].getParent());
    }

    @Override
    public List<String> getHeaderFiles() {
        File[] jnis = findJNIHeaders();
        return Arrays.asList("<" + jnis[0] + ">", "<" + jnis[1] + ">");
    }

    private static File[] findJNIHeaders() throws IllegalStateException {
        final File jreHome = new File(System.getProperty("java.home"));
        final File include = new File(jreHome.getParentFile(), "include");
        final File[] jnis = {
            new File(include, "jni.h"),
            new File(new File(include, "linux"), "jni_md.h"),
        };
        return jnis;
    }
}

The good thing is that jni.h is inside of every JDK, so one can use the java.home property to locate the necessary header files. The actual logic can, of course, be made more robust and OS-independent.

Implementing any JVM native method in Java and/or making callbacks to the JVM with Substrate VM should now be as easy as expanding upon the given example and invoking native-image.