Jaded - Java Deserialization for Rust

Java has a much maligned (for good reason) serialization system built into the standard library. The output is a binary stream mapping the full object hierarchy and the relations between them.

The stream also includes definitions of classes and their hierarchies (super classes etc). The full specification is defined here.

In any new application there are probably better ways to serialize data with fewer security risks but there are cases where a legacy application is writing stuff out and we want to read it in again. If we want to read it in a separate application it'd be good if we weren't bound to Java.

Example

In Java

import java.io.FileOutputStream;
import java.io.ObjectOutputStream;
import java.io.Serializable;
public class Demo implements Serializable {
    private static final long serialVersionUID = 1L;
    private String message;
    private int i;
    public Demo(String message, int count) {
        this.message = message;
        this.i = count;
    }
    public static void main(String[] args) throws Exception {
        Demo d = new Demo("helloWorld", 42);
        try (FileOutputStream fos = new FileOutputStream("demo.obj", false);
                ObjectOutputStream oos = new ObjectOutputStream(fos);) {
            oos.writeObject(d);
        }
    }
}

From Rust

use std::fs::File;
use jaded::{Parser, Result};

fn main() -> Result<()> {
    let sample = File::open("demo.obj").expect("File missing");
    let mut parser = Parser::new(sample)?;
    println!("Read Object: {:#?}", parser.read()?);
    Ok(())
}

Output from Rust

Read Object: Object(
    Object(
        ObjectData {
            class: "Demo",
            fields: {
                "i": Primitive(
                    Int(
                        42,
                    ),
                ),
                "message": JavaString(
                    "helloWorld",
                ),
            },
            annotations: [],
        },
    ),
)

Conversion to Rust types

For most uses cases, the raw object representation is not very ergonomic to work with. For ease of use, types can implement FromJava, and can then be read directly from the stream.

In the majority of cases this implementation can be automatically derived by enabling the derive feature.

#[derive(Debug, FromJava)]
struct Demo {
    message: String,
    i: i32,
}

Demo objects can then be read directly by the parser

fn main() -> Result<()> {
    let sample = File::open("demo.obj").expect("File missing");
    let mut parser = Parser::new(sample)?;
    let demo: Demo = parser.read_as()?;
    println!("Read Object: {:#?}", demo);
    Ok(())
}

Output from rust

Read Object: Demo {
    message: "helloWorld",
    i: 42,
}

Objects with custom writeObject methods

Often classes, including many in the standard library, customise the way they are written using a writeObject method that complements the builtin serialization methods for fields. This data is written as an embedded stream of bytes and/or objects. These cannot be associated with fields without the original Java source so are included in the annotations field of the ObjectData struct (empty in the example above).

As this stream often contains important data from the class, a mechanism is provided to read useful data from it using an interface similar to the ObjectInputStream that would be used in the Java class itself.

An example of custom serialization in Java is the ArrayList. The source for its writeObject methods can be seen here but the gist is that it writes the number of elements it contains, then writes each element in turn.

Because the embedded custom stream could contain anything we have to manually implement the methods to read from it but these can then be used by the derived implementation of FromJava:

In Java

import java.util.List;
import java.util.ArrayList;
import java.io.FileOutputStream;
import java.io.ObjectOutputStream;
public class Demo {
    public static void main(String[] args) throws Exception {
        List<String> keys = new ArrayList<>();
        keys.add("one");
        keys.add("two");
        keys.add("three");
        try (FileOutputStream fos = new FileOutputStream("demo.obj", false);
                ObjectOutputStream oos = new ObjectOutputStream(fos);) {
            oos.writeObject(keys);
        }
    }
}

In rust

use std::fs::File;
use jaded::{Parser, Result, FromJava, FromJava, AnnotationIter, ConversionResult};

#[derive(Debug, FromJava)]
struct ArrayList<T> {
    // Size is written as a 'normal' field
    size: i32,
    // values are written to the custom stream so need attributes
    #[jaded(extract(read_values))]
    values: Vec<T>,
}

// extraction method must be callable as
//     function(&mut AnnotationIter) -> ConversionResult<S> where S: Into<T>
// Where T is the type of the field being assigned to.
fn read_values<T>(annotations: &mut AnnotationIter) -> ConversionResult<Vec<T>>
where
    T: FromJava
{
    (0..annotations.read_i32()?)
        .into_iter()
        .map(|_| annotations.read_object_as())
        .collect()
}


fn main() -> Result<()> {
    let sample = File::open("demo.obj").expect("File missing");
    let mut parser = Parser::new(sample)?;
    let array: ArrayList<String> = parser.read_as()?;
    println!("{:#?}", array);
    Ok(())
}

This gives the array list as expected

ArrayList {
    size: 3,
    values: [
        "one",
        "two",
        "three",
    ],
}

FromJava is implemented for Option<T> and Box<T> so that recursive structs can be deserialized and null fields in the serialized class can be handled. Note that if a field is null the conversion will fail unless that field is given as Option<T>. The example above would have failed if there was a null string in the serialized list. Changing values to be Vec<Option<T>> would allow it to still be read.

Renaming fields

In Java conventions, field names use camelCase whereas Rust field names use snake_case. By default, the derive macro looks for a field named the same as the mapped field in Rust so to prevent Rust structs needing to use camelCase, fields can be given attributes to use a different field in the Java class.

#[derive(FromJava)]
struct Demo {
    #[jaded(field = "fooBar")]
    foo_bar: String,
}

If all fields are to be renamed, the struct can be given a 'rename' attribute. This will convert all field names to camelCase before reading them from Java. Individual fields can still be overridden if required.

#[derive(FromJava)]
#[jaded(rename)]
struct Demo {
    foo_bar: String,
}

Polymorphism

In Java, a field can be declared as an interface and the concrete implementation cannot be known until runtime. Jaded can go some way towards deserializing these fields, using the built in derive macro with an enum.

Each variant of the enum can be assigned a concrete implementation and the fully qualified class name (FQCN) of the object being read will determine which variant is returned.

For instance, to read a field declared as a list, you might define an enum as follows

#[derive(FromJava)]
enum List<T> {
    #[jaded(class = "java.util.ArrayList")] //  without generics
    ArrayList(
        #[extract(read_list)] // See above for read method
        Vec<T>
    ),
    #[jaded(class = "java.util.Collections$EmptyList")]
    Empty,
    #[jaded(class = "java.util.Arrays$ArrayList")]
    // result of using Arrays.asList in Java
    Array {
        a: Vec<T>, // Array is written to field called a
    },
}

Combined with the from field attribute and a From<List> implementation, this enables fields declared as List in Java to be read to Vec<T> in rust.

While this helps to support polymorhpism, it still requires all potential implementations to be known up front. For many use cases this should be adequate.

Features

derive

Allow FromJava to be derived automatically

serde

Add serde serialize/deserialize support for the intermediate types (PrimitiveType, ObjectData, Value and Content). This does not support using serde annotations for deserialisation to user types - FromJava is still required for that - only that raw data can be written to other formats without knowing the data type beforehand.

Limitations

Java Polymorphism

In Java, a field can be declared as an interface and the concrete implementation can be anything. This means that in Rust we can't reliably convert read objects to structs unless we know that a stream is going to be using a specific implementation. While deserializing to an enum would cover most of the common cases, there is nothing stopping some client code creating a CustomList with a completely different serialized representation and using that in the class that is being read in Rust.

Ambiguous serialization

Unfortunately, there are also limits to what we can do without the original code that created the serial byte stream. The protocol linked above lists four types of object. One of which, classes that implement java.lang.Externalizable and use PROTOCOL_VERSION_1 (not been the default since v1.2), are not readable by anything other than the class that wrote them as their data is nothing more than a stream of bytes.

Of the remaining three types we can only reliably deserialize two.

• 'Normal' classes that implement java.lang.Serializable without having a writeObject method

  These can be read as shown above

• Classes that implement Externalizable and use the newer PROTOCOL_VERSION_2

  These can be read, although their data is held fully by the annotations fields of the ObjectData struct and the get_field method only returns None.

• Serializable classes that implement writeObject

  These objects are more difficult. The spec above suggests that they have their fields written as 'normal' classes and then have optional annotations written afterwards. In practice this is not the case and the fields are only written if the class calls defaultWriteObject as the first call in their writeObject method. This is mentioned as a requirement in the spec so we can assume that this is correct for classes in the standard library but it is something to be aware of if user classes are being deserialized.

A consequence of this is that once we have found a class that we can't read, it is difficult to get back on track as it requires picking out the marker signifying the start of the next object from the sea of custom data.

Future plans

• Add implementations of FromJava for common Java and Rust types so that for instance ArrayList and HashMap can be read to the equivalent Vec and HashMap types in Rust.
• Possible tie in with Serde. I've not yet looked into how the serde data model works but this seems like it would be a useful way of accessing Java data.
• Reduce the amount of cloning of data required. As the Java stream contains back references to previous objects in the stream, the actual data in objects read, can't be passed to the caller. Currently, calling read on a parser instance builds the next object by cloning any data referenced by the next object and returning that. This can lead to the same data being cloned many times. It would be better to keep an internal pool of read objects and return objects built of references to that pool. This would mean that read could not be called while a reference to a returned object was still held but if required, the clone could be made on the client side.

State of development

Very much a work in progress at the moment. I am writing this for another application I am working on so I imagine there will be many changes in the functionality and API at least in the short term as the requirements become apparent. As things settle down I hope things will become more stable.

Contributions

As this project it is still very much in a pre-alpha state, I imagine things being quite unstable for a while. That said, if you notice anything obviously broken or have a feature that you think would be useful that I've missed entirely, do open issues. I'd avoid opening PRs until it's been discussed in an issue as the current repo state may lag behind development.