JarCastingIntro
Jaccall makes C libraries accessible from Java without the need to write any native code. It is project similar to JNA or BridJ.
Status:
- available on maven central: 1.0.5
- All major features implemented. In maintenance mode.
- Feature complete [ Linker API, Pointer API, Struct API, Function Pointer API ]
Jaccall's does not try to be Java, but instead tries to make C accessible in Java.
- What you allocate, you must free yourself. watch out for memory leaks!
- Cast to and from anything to anything. Watch out for cast mismatches!
- Read and write to and from anything to anything. Watch out for segfaults!
Design goals:
- Simple usage.
- Simple runtime API.
- No config files.
- C only.
- Support for Linux: aarch64, armv7hf, armv7sf, armv6hf, x86_64, i686.
- Support for all common use cases: unions, callbacks, pointer-to-pointer, ...
Comparison with other libraries
Jaccall was born out of a frustration with existing solutions. Existing solutions have the nasty trade-off of having a complete but cumbersome API and slow runtime, or have excellent speed and good API but suffer from scope creep while lacking armhf support.
Jaccall tries to remedy this by strictly adhering to the KISS princicple.
Overview
Linker API
The linker API forms the basis of all native method invocation. Without it, you wouldn't be able to call any native methods.
To call a C method, we must create a Java class where we define what C method we are interested in, what they look like and where they can be found. This is done by mapping Java methods to C methods, and providing additional information through annotations.
A linker example
C
library: libsomething.so
header: some_header.h
struct test {
char field0;
unsigned short field1;
int field2[3];
int *field3;
};
...
struct test do_something(struct test* tst,
char field0,
unsigned short field1,
int* field2,
int* field3);
Java SomeHeader.java
@Lib("something")
public class SomeHeader {
static {
Linker.link(SomeHeader.class);
}
@ByVal(StructTest.class)
public native long do_something(@Ptr(StructTest.class) long tst,
byte field0,
@Unsigned short field1,
@Ptr(int.class) long field2,
@Ptr(int.class) long field3);
}
This Java class exposes the C header file some_header.h to the Java side and informs the linker where these symbols (methods) can be resolved. This is done by providing the @Lib(...) annotation who's value must match the name part of libsomething.so. This whole flow is triggered by calling Linker.link(...).
Don't worry about the struct, that is handled in the Struct API.
In order to pass data back and forth between Java and C, there are a few mapping rules to keep in mind.
- Java method name must match C method name
- Java method must be declared
native - Java method must be declared
public - Java method must only consist of a specific set of primitives for both arguments and return type.
Mapping
The Java mapping tries to match it's C counterpart as close as possible. There are however a few non intuitive exceptions. Let's have a look on how C types map to their Java counterpart.
| C | Java |
|---|---|
| unsigned char or char | byte |
| unsigned char | @Unsigned byte |
| short | short |
| unsigned short | @Unsigned short |
| int | int |
| unsigned int | @Unsigned int |
| long | long |
| unsigned long | @Unsigned long |
| long long | @Lng long |
| unsigned long long | @Unsigned @Lng long |
| float | float |
| double | double |
| struct foo | @ByVal(Foo.class) long |
| union bar | @ByVal(Bar.class) long |
| foo* | @Ptr(Foo.class) long |
The Java primitive types boolean and char do not have a corresponding C type and are not allowed.
The class argument for @Ptr is optional.
Arrays should be mapped as a pointer of the same type.
By value By reference
Java does not support the notion of passing by reference or by value. By default, all method arguments are passed by value in Java, inlcuding POJOs which are actually pointers internally. This limits the size of a single argument in Java to 64-bit. As such, Jaccall can not pass or return a C struct by value. Jaccall works around this problem by allocating heap memory and copyin/reading struct-by-value data. Jaccall must then only pass a pointer between Java and the native side.
The drawback of this approach is that all returned struct-by-value data must be freed manually!
Internals
Jaccall has a compile time step to perform both fail-fast compile time checks and Java source code generation. To aid the Linker in processing a natively mapped Java class, the LinkerGenerator does an initial pass over any @Lib annotated class to verify it's integrity and mapping rules. If this succeeds, it does a second pass and generates a Foo_Jaccall_LinkSymbols.java soure file for every Foo.java annotated with @Lib. This file should not be used by application code. This file contains linker data to aid the Linker in linking the required native Java methods to it's C counterpart.
For every mapped method, 4 parts of linker data are generated.
- The method name (the C symbol name).
- The number of arguments.
- A LibFFI call interface.
- A JNI signature.
If we reiterate our first mapping example
C some_header.h
struct test {
char field0;
unsigned short field1;
int field2[3];
int *field3;
};
...
struct test do_something(struct test* tst,
char field0,
unsigned short field1,
int* field2,
int* field3);
Java SomeHeader.java
@ByVal(StructTest.class)
public native long do_something(@Ptr(StructTest.class) long tst,
byte field0,
short field1,
@Ptr(int.class) long field2,
@Ptr(int.class) long field3);
The generated linker data for this mapping: SomeHeader_Jaccall_LinkSymbols.java
...
@Generated("org.freedesktop.jaccall.compiletime.LinkerGenerator")
public final class SomeHeader_Jaccall_LinkSymbols extends LinkSymbols {
public SomeHeader_Jaccall_LinkSymbols() {
super(new String[]{"do_something"},
new byte[]{5},
new long[]{JNI.ffi_callInterface(StructTest.FFI_TYPE,
JNI.FFI_TYPE_POINTER,
JNI.FFI_TYPE_SINT8,
JNI.FFI_TYPE_UINT16,
JNI.FFI_TYPE_POINTER,
JNI.FFI_TYPE_POINTER)},
new String[]{"(JBSJJ)J"});
}
}
"do_something"The name of the method5The number of argumentsJNI.ffi_callInterface(...)The libffi call interface."(JBSJJ)J"The JNI method signature.
Linker data of different methods matches on array index.
Pointer API
A pointer example
C
...
size_t int_size = sizeof(int);
void* void_p = malloc(int_size);
int* int_p = (int*) void_p;
...
free(int_p);
This example is pretty self explenatory. A new block of memory is allocated of size 'int'. This block of memory is then cast to a pointer of type int, and finally the memory is cleaned up.
Using Jaccall this translates to
//(Optional) Define a static import of the Pointer and Size classes
//to avoid prefixing all static method calls.
import static Pointer.*
import static Size.*
...
//Calculate the size of Java type `Integer` wich corresponds to a C int.
int int_size = sizeof((Integer)null);
//Allocate a new block of memory, using `int_size` as it's size.
Pointer<Void> void_p = malloc(int_size);
//Do a pointer cast of the void pointer `void_p` to a pointer of type `Integer`.
Pointer<Integer> int_p = void_p.castp(Integer.class);
...
//free the memory pointed to by `int_p`
int_p.close();
Stack vs Heap
C has the concept of stack and heap allocated memory. Unfortunately this doesn't translate well in Java. Jaccall tries to alleviate this by utilizing Java's garbage collector mechanism. To understand the idea behind this, it's important to know the difference between memory allocated with Pointer.malloc(..) and Pointer.nref(..).
A pointer that refers to malloc memory is not subject to Java's garbage collector and will never be freed manualy. Just like in C, the program is expected to free the memory when it's done with it.
A pointer that refers to nref memory is subject to Java's garbage collector and as such requires no manual call to free. It is meant to mimic C's stack allocated memory. It is strongly advices to use this memory solely within the scope of the method that called nref.
One can also specifically scope heap allocated memory as Jacall defines a Pointer<...> as an AutoClosable. Using Java's try-with-resource concept, we can precisely scope heap allocated memory.
Do not use try-with-resource on nref allocated memory, as it will result in a segfault caused by a double free.
C
int some_int = 5;
int* int_p = &some_int;
...
//`int_p` becomes invalid once method ends
Using Jaccall nref this translates to
//(Optional) Define a static import of the Pointer class to avoid prefixing all static method calls.
import static Pointer.*
...
//define an integer
int some_int = 5;
//allocate a new block of scoped memory with `some_int` as it's value.
Pointer<Integer> int_p = nref(some_int);
...
//`int_p` becomes invalid once Java's garbage collector kicks in.
Using Jaccall malloc this translates to
//(Optional) Define a static import of the Pointer class to avoid prefixing all static method calls.
import static Pointer.*
import static Size.*
...
//define an integer
int some_int = 5;
//allocate a new block of scoped memory with `some_int` as it's value.
try(Pointer<Integer> int_p = malloc(sizeof(some_int).castp(Integer.class)){
int_p.write(some_int);
...
}
...
//`int_p` becomes invalid once try block ends.
There are some notable differences between the C and Java example. In the C example, only one block of memory is used to define some_int, int_p is simply a reference to this memory. This block of memory is method scoped (stack allocated). Once the method exits, the memory is cleaned up.
On the Java side however things are a bit different. A Java object (primitive) is defined as some_int. Next a new block of memory int_p is allocated on the heap, and the value of some_int is copied into it. This operation is reflected in the call Pointer.write(some_int). Because we defined int_p inside a try-with-resources, it will be freed automatically with a call to close() once the try block ends.
It is important to notice that there is nothing special about Pointer.nref(some_int). It's merely a shortcut for
Pointer.malloc(Size.sizeof(some_int)).castp(Integer.class).write(some_int);
but unlike a plain malloc has it's lifecycle tracked by the Java garbage collector.
Memory read write
Let's extend our first basic example and add some read and write operations.
C
...
size_t int_size = sizeof(int);
void* void_p = malloc(int_size);
int* int_p = (int*) void_p;
*int_p = 5;
int int_value = *int_p;
...
free(int_p);
The equivalent Java code:
import static Pointer.*
import static Size.*
...
int int_size = sizeof((Integer)null);
Pointer<Void> void_p = malloc(int_size);
Pointer<Integer> int_p = void_p.castp(Integer.class);
//write an int with value 5 to memory
int_p.write(5);
//read (dereference the pointer) an int from memory
int int_value = int_p.dref();
...
int_p.close();
The data that can be written and read from a pointer object in Jaccall depends on data type it refers to. This is why it's necessary that we perform a pointer cast using castp(Integer.class). This creates a new pointer object that can read and write integers.
There are 3 different cast operations that can be performed on a pointer object.
- an ordinary cast, using
cast(Class<?>). Cast a pointer to any primitive or struct type. - a pointer cast, using
castp(Class<?>). Cast a pointer to a pointer of another type. - a pointer to pointer cast, using
castpp(). Cast a pointer to a pointer-to-pointer.
Starting from our basic example
import static Pointer.*
import static Size.*
...
int int_size = sizeof((Integer)null);
Pointer<Void> void_p = malloc(int_size);
//Perform a pointer cast, the resulting pointer can be used to read and write an integer.
Pointer<Integer> int_p = void_p.castp(Integer.class);
int_p.write(5);
int int_value = int_p.dref();
//Perform a pointer to pointer cast.
Pointer<Pointer<Integer>> int_pp = int_p.castpp();
//Dereferencing will cause a pointer object to be created with address 5, or possibly even segfault on a 64-bit system!
Pointer<Integer> bad_int_p = int_pp.dref();
//Perform an ordinary cast, `some_long` will now contain the address of our `int_p` pointer!
long some_long = int_p.cast(Long.class);
...
int_p.close();
In most cases, an ordinary cast using cast(Class<?>) will not be needed.
Beware that when casting to a pointer-to-pointer using castp(Pointer.class), you will end up with a Pointer<Pointer<?>> object. You will be able to dereference the Pointer<Pointer<?>> object to the underlying Pointer<?>, but you will not be able to write or dereference this resulting pointer as Jaccall does not know what type it should refer too. Internally Jaccall will represent the Pointer<?> object as a Pointer<Void>.
It might not be immediatly obvious at first but using castp(Class<?>) and castpp() we can cast any pointer to any other type of pointer-to-pointer-to-pointer ad infinitum.
An example where we receive an address from a jni library. We know the address represents a char***, and as such want to create a Pointer<Pointer<Pointer<Byte>>>.
//get a native address from a jni library.
long some_native_address = ...;
//wrap the address in a untyped pointer
Pointer<Void> void_p = Pointer.wrap(some_native_address);
//cast to a byte pointer
Pointer<Byte> byte_p = void_p.castp(Byte.class);
//cast to a pointer-to-pointer
Pointer<Pointer<Byte>> byte_pp = byte_p.castpp();
//cast to a pointer-to-pointer-to-pointer
Pointer<Pointer<Pointer<Byte>>> byte_ppp = byte_pp.castpp();
We can rewrite the above example more briefly
import static Pointer.*
...
long some_native_address = ...;
//wrap in a pointer-to-pointer-to-char pointer
Pointer<Pointer<Pointer<Byte>>> byte_ppp = wrap(Byte.class,some_native_address).castpp().castpp();
Arrays
Up until now we've worked with single element pointers. Because a C array can be represented as a pointer, accessing a C array is surprisingly easy. Our pointer object knows the size of the type it's refering to, which in turn makes indexed read and writes straightforward.
We reiterate our previous read/write example, only this time we allocate space for multiple integers.
import static Pointer.*;
import static Size.*;
...
int int_size = sizeof((Integer)null);
//allocate space for 3 integers
Pointer<Void> void_p = malloc(int_size*3);
//cast to an ordinary int pointer
Pointer<Integer> int_p = void_p.castp(Integer.class);
//write an int with value 5 to index 0.
int_p.writei(0,5);
//write an int with value 6 to index 1.
int_p.writei(1,6);
//write an int with value 7 to index 2.
int_p.writei(2,6);
//dereference the pointer at index 0
int int_value_0 = int_p.dref(0);
//dereference the pointer at index 1
int int_value_1 = int_p.dref(1);
//dereference the pointer at index 2
int int_value_0 = int_p.dref(2);
...
int_p.close();
We can also use Pointer.nref(...) to allocate an array.
import static Pointer.*;
...
//nref uses a vararg parameter, so we can use it with both arrays and classic function arguments
Pointer<Integer> int_p_varargs = nref(1,2,3,4,5);
Integer[] array = {1,2,3,4,5);
Pointer<Integer> int_p_array = nref(array);
Address manipulation
In C, one can read and change the actual address value of a pointer. In Jaccall this is no different. The pointer object exposes it's address either directly through an address field of type long, or it can be casted to a long.
Pointer<Void> void_pointer = ...
//`void_pointer_adr` now contains the actual address of `void_pointer`
long void_pointer_adr = void_pointer.address;
//`void_pointer_adr_cast` now contains exactly the same value as `void_pointer_adr`.
long void_pointer_adr_cast = void_pointer.cast(Long.class)'
We can also easily offset a pointer address while keeping the type it points to.
Pointer<String> char_pointer = ...
//`char_pointer_offset` address is now incremented by 2 compared to `char_pointer` address.
Pointer<String> char_pointer_offset = char_pointer.offset(2);
Pointer types
A pointer object only supports a limited amount of Java types it can refer to. This is because it has to perform a mapping operation from the underlying C type to the equivalent Java type.
Following types are supported
| C | Java |
|---|---|
| void | Void or void |
| char | Byte or byte |
| unsigned short or short | Short or short |
| unsgined int or int | Integer or int |
| unsigned long or long | CLong |
| unsigned long long or long long | Long or long |
| float | Float or float |
| double | Double or double |
| foo* | Pointer |
| char* | String |
Java primitives like boolean (Boolean) or char (Character) are not supported for the simple reason that they do not have a good C counterpart. A boolean type does not exist in C, and a Java char is actually an unsigned 16-bit integer that is used as an utf-16 character as opposed to C's 8-bit char type.
Struct API
A struct example
Jaccall allows you to map any struct or union type in Java. Let's have a look at our previous example that contained a struct definition:
C
struct test {
char field0;
unsigned short field1;
int field2[3];
int *field3;
};
...
Mapping this struct in Java using Jaccall
...
import static CType.CHAR;
import static CType.INT;
import static CType.POINTER;
import static CType.UNSIGNED_SHORT;
...
@Struct(value = {
@Field(type = CHAR,
name = "field0"),
@Field(type = UNSIGNED_SHORT,
name = "field1") ,
@Field(type = INT,
cardinality = 3,
name = "field2"),
@Field(type = POINTER,
dataType = int.class,
name = "field3")
})
public class Test extends Test_Jaccall_StructType {
}
Mapping a union is completely analogue. C some_header.h
union test {
char field0;
int field1;
};
...
In Java this becomes
...
import static CType.CHAR;
import static CType.INT;
...
@Struct(value = {
union = true,
@Field(type = CHAR,
name = "field0"),
@Field(type = INT,
name = "field1"),
})
public class Test extends Test_Jaccall_StructType {
}
The @Struct annotation defines the layout of the native C struct in Java. This layout is parsed during compilation to generate accessor code. This generated code is put in a Java class with name Foo_Jaccall_StructType, where Foo is the name of the class that has the @Struct annotation. To use this accessor code, simply extend the generated class. In our exmaple this becomes extends Test_Jaccall_StructType.
The generated accessor class is part of the internal Jaccall API and should never be used directly. Instead always inherit from it.
The following rules apply when annotating a class with @Struct.
- A class annotated with
@Structmust have a default no-arg constructor. - A class annotated with
@Structmust not be abstract. - A class annotated with
@Structmust have at least one@Field. - A class annotated with
@Structmust extend the equivalent generated accessor class. - A class annotated with
@Structmust have unique@Fieldnames. - A class annotated with
@Structmust not have a static field with nameSIZE. - A class annotated with
@Structmust be public. - A class annotated with
@Structmust be a class. - A class annotated with
@Structmust be a top level class.
Field definitions
TODO
Usage
Using a Jaccall struct in Java is very similar as how you would use a C struct. You can create a new one directly
C
struct test testStruct;
testStruct.field0 = (char)123;
...
int field1 = testStruct.field1;
Java
Test testStruct = new Test();
testStruct.field0((byte)123);
...
int field1 = testStruct.field1();
or you can allocate a block of memory first, and map it as a struct.
C
void* voidPointer = malloc(sizeof(struct test));
struct test *testPointer = (struct test *) voidPointer;
struct test testStruct = *testPointer;
testStruct.field0 = (char)123;
...
int field1 = testStruct.field1;
Java
Pointer<Void> voidPointer = Pointer.malloc(Test.SIZE);
Pointer<Test> testPointer = voidPointer.pcast(Test.class);
Test testStruct = testPointer.dref();
testStruct.field0((byte)123);
...
int field1 = testStruct.field1();
The important difference between the 2 cases is that the first one creates a struct on the stack (for C), which translated to memory subject to garbage collection in Java. While in the second case the memory has to be mannually freed.
To get a pointer to a 'stack' allocated struct, use Pointer.ref(..). This will consistently return the same address of the struct.
Ref vs Nref
In our previous (non struct) examples we saw that a call to nref would allocated a new block of managed memory and return it's address. This in contrast with ref, which consistenly returns the same memory address and only accepts a struct. The important difference to notice here is that nref is short for 'new reference', while ref simply means 'reference', as the 'new' already happened with a call to new FooStruct().
The important thing to remember here is that in the case of nref the memory lifecycle is determined by the pointer object that was returned. In case of ref the memory lifecycle is determined by struct object itself.
Function Pointer API
A function pointer example
C
typedef char(*testFunc)(struct test*, unsigned int, struct test);
Java
@Functor
public interface TestFunc {
byte $(@Ptr(TestStruct.class) long arg0, @Unsigned int arg1, @ByVal(TestStruct.class) long arg2);
}
Defining a single method interface as a @Functor will signal Jaccall to generate Java stubs as well as a factory for instantiating those stubs. The rules for mapping a function pointer are the same as for mapping a native method in the Linker API, in addition the method name MUST be defined as $. This allows a user to go from a C pointer to a callable Java object, as well as defining a Java method and expose it as a C function pointer!
The factory is generated in the same package as the annoated interface, and will have it's named derived from the declared interface: Pointer+<interface name>. The factory for our TestFunc will thus be named PointerTestFunc.
The generated factory can be used like this.
When invoking a function pointer from C:
//get an address from C through jni or jacall.
long c_func_ptr = ...;
//wrap the address so we can invoke it on the Java side.
PointerTestFunc cFuncPointer = PointerTestFunc.wrapFunc(c_func_ptr);
...
//invoke the function, this will execute the actual C function.
cFuncPointer.$(arg0, arg1, arg2);
...
//PointerTestFunc extends from the Pointer class, and as such, is a pointer itself.
assert(cFuncPointer.address == c_func_ptr);
When constructing a function pointer from Java.
//the implementation in Java of a C function
public byte javaFunction(@Ptr(StructType.class) final long arg0, @Unsigned final int arg1, @ByVal(TestStruct.class) final long arg2) {
...
}
...
//create a new function pointer that points to the above java function.
final PointerTestFunc pointerTestFunc = PointerTestFunc.nref(new TestFunc() {
@Override
public byte $(@Ptr final long arg0, @Unsigned final int arg1, @ByVal(TestStruct.class) final long arg2) {
return javaFunction(arg0, arg1, arg2);
}
});
//or more briefly (Java 8)
final PointerTestFunc pointerTestFunc = PointerTestFunc.nref(this::javaFunction);
...
//pass on the function pointer address to the native side
someNativeFunction(pointerTestFunc.address);
...
//we can also execute the function pointer in Java
pointerTestFunc.$(arg0, arg1, arg2);
It's important to notice that a a call to nref in this case will not (yet) automatically reclaim memory once the functor object goes out of scope. As such it is a source of potential memory leaks which will be addressed in a future release of Jaccall.
