Thanks Sean,
No I hadn't seen it, I've just read it, will probably need to read it
again to appreciate it fully...
It certainly identifies all the issues I'm aware of, as well as being
respectful of the original implementors (many of whom participated in
Apache River when Jini was donated to Apache), I came to the same
conclusion with circular object graphs; the benefits don't outweigh the
cost.
We also use annotations instead of interfaces,to annotate the class and
constructor, so that overriding classes don't automagically inherit the
functionality.
At this time, we haven't reimplemented deconstruction, we are using
ObjectOutputStream with serializers, which are basically serialization
proxy's for existing classes, we have fully reimplemented
deserialization using constructors.
Agree with serial from being independant of the wire protocol, so any
serialization scheme can be used, this is an excellent idea of course.
The constructors / deconstructors have identified that serial form is
really just a parameter list. Developers will want to make defensive
copies of mutable state, just like public api methods.
We did consider constructors with multiple parameters, but decided
against it for the following reasons:
1. We didn't care about parameter order (tuples), or the order in
which they were serialized / deserialized, we only cared about
parameter names and types.
2. For encapsulation we didn't want subclasses having to manage the
serial form of superclasses, we wanted them to remain as
independant as possible, so they don't inadvertantly break.
* For example, a library superclass adds a serial form
parameter, or changes a type, in its serial form. The
child class would have to be aware of the changes in order
to pass the correct parameters to the correct superclass
constructor.
* Different serial version constructors would result in the
loss of later version superclass state when child classes
call an earlier version.
3. We settled on a caller sensitive parameter that is passed to the
deserialization constructor.
* Encapsulation: Each class in an inheritance heirarch only
has access to it's own serial form.
* The serial form of each class is independant and may evolve
independantly.
* Each class in the inheritance heirarchy is responsible for
checking it's own invariants, including the ability to
create superclass instances, even if a superclass is
abstract for checking inter class invariants.
4. It was less work for the framework to populate a standard
parameter object, with serial form, the framework didn't need to
worry about inspecting the constructor signature and determining
the parameter order.
5. One constructor could be used for different versions.
6. We currently use |serialPersistentFields to declare serial form,
but there is probably a better way of doing this, perhaps a way
that also documents different serial form versions.|
Regards,
Peter.
On 20/08/2019 7:55 AM, Sean Mullan wrote:
Brian Goetz (copied) has done a lot of thinking in the serialization
area, so I have copied him. Not sure if you have seen it but he
recently posted a document about some of his ideas and possible future
directions for serialization:
http://cr.openjdk.java.net/~briangoetz/amber/serialization.html
--Sean
On 8/17/19 10:22 PM, Peter Firmstone wrote:
Thanks Sean,
You've gone to some trouble to answer my question, which demonstrates
you have considered it.
I donate some time to help maintain Apache River, derived from Sun's
Jini. Once Jini depended on RMI, today, not so much, it still has
some dependencies on some RMI interfaces, but doesn't utilise JRMP
although it provides some backward compatibilty enable it.
But my point is, we heavily utilise java Serialization, and have an
independant implementation of a subset of Java Serialization
(originating from Apache Harmony). We do this for security as we use
an annotated serialization constructor. Serial form is unchanged,
we have Serializers for commonly used java library objects, for
example, we have a "PermissionSerializer", but we don't have a
"PermissionCollectionSerializer" or "PermissionsSerializer" (for
java.security.Permissions). Incidentally, we have found we do not
need the ability to serialize circular object graphs. Throwable is
an object that has a circular object graph, but that circular object
graph can be linked up after deserialization.
Permission implementing Serializable is probably not too much of a
threat, as these objects are effectively immutable after lazy
initialization.
ProtectionDomain calls java.security.Permissions::setReadOnly during
it's construction.
ProtectionDomain::getPermissions returns internal
java.security.Permissions. If this is serialized, then the readOnly
internal state can be written to as the internal object references
are accessible from within the stream.
Admitedly, the attacker would already need to have some privilege, to
have access to a ProtectionDomain, so it's a path of privilege
escallation. I'm not talking about gadget attacks and
deserialization of untrusted data, I'm talking about breaking
encapsulation.
Even though we are heavily dependant on Java Serialization, we are
very careful when we implement it, and avoid implementing it when
possible. Hindsight is 20:20, but given we are now seeing some Java
SE backward compatibility breakages, perhaps it might be worth
considering breaking serialization. I don't mean we need to
necessarily break object serial form, but making the Java
serialization API explicit with subset of existing api features, that
makes long term maintenace and security less of a burden and removing
support for Serialization of some objects, where it is seldom used,
perhaps using a JEP that requests developers to consider which
library objects actually need to be serializable.
Something we do in our Java Serialization API is require that mutable
deserialized objects are defensively copied during object
construction (serial fields are deserialized before an object is
constructed, the deserialized fields are accessible via a parameter
passed in during construction. We have tools that assist developers
to check deserialized Java Collections contain the expected object
types for example, so during object construction the developer has to
replace the Collection with a new instance and copy the contents to
the new Collection after checking the type of each object contained
therein. Also we don't actually serialize Java Collections, we have
standard serial forms for List, Set and Map, so these serial forms
are equal, similar to the List, Set and Map contracts. By doing
this, Collections don't actually need to implement Serializable at
all, as a Serializer becomes responsible for their serialization.
This also means that all Collections must be accessed by interfaces,
rather than implementation classes, so the deserialization
constructor, must defensively copy them into their preferred
Collection instance. It's a bit like dependency injection.
I know it would take time, and there would be some pain, but long
term it would save a lot of maintenance developer time.
Regards,
Peter.
On 17/08/2019 12:50 AM, Sean Mullan wrote:
On 8/15/19 8:18 PM, Peter Firmstone wrote:
Hi Roger,
+1 for writeReplace
Personally I'd like to see some security classes break backward
compatibility and remove support for serialization as it allows
someone to get references to internal objects, especially since
these classes are cached by the JVM. Which makes
PermissionCollection.setReadOnly() very easy to bypass, by adding
permissions to internal collections once you have a reference to them.
Does anyone have any use cases for serializing these objects?
These objects are easy to re-create by sending or recieving and
parsing strings, because they are built from text based policy
files, and when you do that, you are validating input, so I never
did fully understand why they were made serializable.
This is briefly explained on page 61 in the "Inside Java 2 Platform
Security" book [1]:
"The Permission class implements two interfaces: java.security.Guard
and java.io.Serializable. For the latter, the intention is that
Permission objects may be transported to remote machines, such as
via Remote Method Invocation (RMI), and thus a Serializable
representation is useful."
The Permission class was introduced in Java SE 1.2 so there were
different motivations back then :)
--Sean
[1] https://www.oracle.com/technetwork/java/javaee/index-141918.html