Apache Commons Logging – Technology Guide

Overview

Overview
- Contents
- Introduction
A Short Introduction to Class Loading and Class Loaders
A Short Guide To Hierarchical Class Loading
A Short Theory Guide To JCL

Introduction

This guide is aimed at describing the technologies that JCL developers and expert users (and users who need to become experts) should be familiar with. The aim is to give an understanding whilst being precise but brief. Details which are not relevant for JCL have been suppressed. References have been included.

These topics are a little difficult and it's easy for even experienced developers to make mistakes. We need you to help us get it right! Please submit corrections, comments, additional references and requests for clarification by either:

posting to the Apache Commons dev mailing list or
creating an issue in JIRA.

TIA

A Short Introduction to Class Loading and Class Loaders

Preamble

This is intended to present a guide to the process by which Java bytecode uses bytecode in other classes from the perspective of the language and virtual machine specifications. The focus will be on deciding which bytecode will be used (rather than the mechanics of the usage). It focuses on facts and terminology.

The process is recursive: it is therefore difficult to pick a starting point. Sun's documentation starts from the perspective of the startup of a new application. This guide starts from the perspective of an executing application.

During this discussion, please assume that each time that class is mentioned, the comments applied equally well to interfaces.

This document is targeted at Java 1.2 and above.

Resolution Of Symbolic References

(LangSpec 12.3.3) The bytecode representation of a class contains symbolic names for other classes referenced.

In practical development terms: If a class is imported (either explicitly in the list of imports at the top of the source file or implicitly through a fully qualified name in the source code) it is referenced symbolically.

(VMSpec 5.4.3) Resolution of a symbolic reference occurs dynamically at runtime and is carried out by the Java Virtual Machine. Resolution of a symbolic reference requires loading and linking of the new class.

Note: references are not statically resolved at compile time.

Loading

(VMSpec 2.17.2) Loading is the name given to the process by which a binary form of a class is obtained by the Java Virtual Machine. Java classes are always loaded and linked dynamically by the Java Virtual Machine (rather than statically by the compiler).

In practical development terms: This means that the developer has no certain knowledge about the actual bytecode that will be used to execute any external call (one made outside the class). This is determined only at execution time and is affected by the way that the code is deployed.

Linking

(VMSpec 2.17.3) Linking is the name used for combining the binary form of a class into the Java Virtual Machine. This must happen before the class can be used.

(VMSpec 2.17.3) Linking is composed of verification, preparation and resolution (of symbolic references). Flexibility is allowed over the timing of resolution. (Within limit) this may happen at any time after preparation and before that reference is used.

In practical development terms: This means that different JVMs may realize that a reference cannot be resolved at different times during execution. Consequently, the actual behavior cannot be precisely predicted without intimate knowledge of the JVM (on which the bytecode will be executed). This makes it hard to give universal guidance to users.

Loading Classes

(VMSpec 2.17.2) The loading process is performed by a ClassLoader.

(VMSpec 5.3) A classloader may create a class either by delegation or by defining it directly. The classloader that initiates loading of a class is known as the initiating loader. The classloader that defines the class is known as the defining loader.

In practical terms: understanding and appreciating this distinction is crucial when debugging issues concerning classloaders.

Bootstrap Classloader

(VMSPEC 5.3) The bootstrap is the base ClassLoader supplied by the Java Virtual Machine. All others are user (also known as application) ClassLoader instances.

In practical development terms: The System classloader returned by Classloader.getSystemClassLoader() will be either the bootstrap classloader or a direct descendant of the bootstrap classloader. Only when debugging issues concerning the system classloader should there be any need to consider the detailed differences between the bootstrap classloader and the system classloader.

Runtime Package

(VMSpec 5.3) At runtime, a class (or interface) is determined by its fully qualified name and by the classloader that defines it. This is known as the class's runtime package.

(VMSpec 5.4.4) Only classes in the same runtime package are mutually accessible.

In practical development terms: two classes with the same symbolic name can only be used interchangeably if they are defined by the same classloader. A classic symptom indicative of a classloader issue is that two classes with the same fully qualified name are found to be incompatible during a method call. This may happen when a member is expecting an interface which is (seemingly) implemented by a class but the class is in a different runtime package after being defined by a different classloader. This is a fundamental java language security feature.

Loader Used To Resolve A Symbolic Reference

(VMSpec 5.3) The classloader which defines the class (whose reference is being resolved) is the one used to initiate loading of the class referred to.

In practical development terms: This is very important to bear in mind when trying to solve classloader issues. A classic misunderstanding is this: suppose class A defined by classloader C has a symbolic reference to class B and further that when C initiates loading of B, this is delegated to classloader D which defines B. Class B can now only resolve symbols that can be loaded by D, rather than all those which can be loaded by C. This is a classic recipe for classloader problems.

Bibliography

VMSpec The Java Virtual Machine Specification, Second Edition
LangSpec The Java Language Specification, Second Edition

A Short Guide To Hierarchical Class Loading

Delegating Class Loaders

When asked to load a class, a class loader may either define the class itself or delegate. The base ClassLoader class insists that every implementation has a parent class loader. This delegation model therefore naturally forms a tree structure rooted in the bootstrap classloader.

Containers (i.e. applications such as servlet engines or application servers that manage and provide support services for a number of "contained" applications that run inside of them) often use complex trees to allow isolation of different applications running within the container. This is particularly true of J2EE containers.

Parent-First And Child-First Class Loaders

When a classloader is asked to load a class, a question presents itself: should it immediately delegate the loading to its parent (and thus only define those classes not defined by its parent) or should it try to define it first itself (and only delegate to its parent those classes it does not itself define). Classloaders which universally adopt the first approach are termed parent-first and the second child-first.

Note: the term child-first (though commonly used) is misleading. A better term (and one which may be encountered on the mailing list) is parent-last. This more accurately describes the actual process of classloading performed by such a classloader.

Parent-first loading has been the standard mechanism in the JDK class loader, at least since Java 1.2 introduced hierarchical classloaders.

Child-first classloading has the advantage of helping to improve isolation between containers and the applications inside them. If an application uses a library jar that is also used by the container, but the version of the jar used by the two is different, child-first classloading allows the contained application to load its version of the jar without affecting the container.

The ability for a servlet container to offer child-first classloading is made available, as an option, by language in the servlet spec (Section 9.7.2) that allows a container to offer child-first loading with certain restrictions, such as not allowing replacement of java.* or javax.* classes, or the container's implementation classes.

Though child-first and parent-first are not the only strategies possible, they are by far the most common. All other strategies are rare. However, it is not uncommon to be faced with a mixture of parent-first and child-first classloaders within the same hierarchy.

Class ClassLoader

The class loader used to define a class is available programmatically by calling the getClassLoader method on the class in question. This is often known as the class classloader.

Context ClassLoader

Java 1.2 introduces a mechanism which allows code to access classloaders which are not the class classloader or one of its parents. A thread may have a class loader associated with it by its creator for use by code running in the thread when loading resources and classes. This classloader is accessed by the getContextClassLoader method on Thread. It is therefore often known as the context classloader.

Note that the quality and appropriateness of the context classloader depends on the care with which the thread's owner manages it.

The Context Classloader in Container Applications

The Javadoc for Thread.setContextClassLoader emphasizes the setting of the context classloader as an aspect of thread creation. However, in many applications the context classloader is not fixed at thread creation but rather is changed throughout the life of a thread as thread execution moves from one context to another. This usage of the context classloader is particularly important in container applications.

For example, in a hypothetical servlet container, a pool of threads is created to handle HTTP requests. When created these threads have their context classloader set to a classloader that loads container classes. After the thread is assigned to handle a request, container code parses the request and then determines which of the deployed web applications should handle it. Only when the container is about to call code associated with a particular web application (i.e. is about to cross an "application boundary") is the context classloader set to the classloader used to load the web app's classes. When the web application finishes handling the request and the call returns, the context classloader is set back to the container classloader.

In a properly managed container, changes in the context classloader are made when code execution crosses an application boundary. When contained application A is handling a request, the context classloader should be the one used to load A's resources. When application B is handling a request, the context classloader should be B's.

While a contained application is handling a request, it is not unusual for it to call system or library code loaded by the container. For example, a contained application may wish to call a utility function provided by a shared library. This kind of call is considered to be within the "application boundary", so the context classloader remains the contained application's classloader. If the system or library code needs to load classes or other resources only visible to the contained application's classloader, it can use the context classloader to access these resources.

If the context classloader is properly managed, system and library code that can be accessed by multiple applications can not only use it to load application-specific resources, but also can use it to detect which application is making a call and thereby provided services tailored to the caller.

Issues with Context ClassLoaders

In practice, context classloaders vary in quality and issues sometimes arise when using them. The owner of the thread is responsible for setting the classloader. If the context classloader is not set then it will default to the system classloader. Any container doing so will cause difficulties for any code using the context classloader.

The owner is also at liberty to set the classloader as they wish. Containers may set the context classloader so that it is neither a child nor a parent of the classloader that defines the class using that loader. Again, this will cause difficulties.

Introduced in Java J2EE 1.3 is a requirement for vendors to appropriately set the context classloader. Section 6.2.4.8 (1.4 text):

This specification requires that J2EE containers provide a per thread
context class loader for the use of system or library classes in
dynamically loading classes provided by the application.  The EJB
specification requires that all EJB client containers provide a per
thread context class loader for dynamically loading system value classes.
The per thread context class loader is accessed using the Thread method
getContextClassLoader.

The classes used by an application will typically be loaded by a
hierarchy of class loaders. There may be a top level application class
loader, an extension class loader, and so on, down to a system class
loader.  The top level application class loader delegates to the lower
class loaders as needed.  Classes loaded by lower class loaders, such as
portable EJB system value classes, need to be able to discover the top
level application class loader used to dynamically load application
classes.

We require that containers provide a per thread context class loader
that can be used to load top level application classes as described
above.

This specification leaves quite a lot of freedom for vendors. (As well as using unconventional terminology and containing the odd typo.) It is a difficult passage (to say the least).

Reflection And The Context ClassLoader

Reflection cannot bypass restrictions imposed by the java language security model, but, by avoiding symbolic references, reflection can be used to load classes which could not otherwise be loaded. Another ClassLoader can be used to load a class and then reflection used to create an instance.

Recall that the runtime packaging is used to determine accessibility. Reflection cannot be used to avoid basic java security. Therefore, the runtime packaging becomes an issue when attempting to cast classes created by reflection using other class loaders. When using this strategy, various modes of failure are possible when common class references are defined by the different class loaders.

Reflection is often used with the context classloader. In theory, this allows a class defined in a parent classloader to load any class that is loadable by the application. In practice, this only works well when the context classloader is set carefully.

More Information

Articles On Class Loaders And Class Loading
Specific Containers

A Short Theory Guide To JCL

Isolation And The Context Class Loader

JCL takes the view that different context class loader indicate boundaries between applications running in a container environment. Isolation requires that JCL honours these boundaries and therefore allows different isolated applications to configure their logging systems independently.

Log And LogFactory

Performance dictates that symbolic references to these classes are present in the calling application code (reflection would simply be too slow). Therefore, these classes must be loadable by the classloader that loads the application code.

Log Implementations

Performance dictates that symbolic references to the logging systems are present in the implementation classes (again, reflection would simply be too slow). So, for an implementation to be able to function, it is necessary for the logging system to be loadable by the classloader that defines the implementing class.

Using Reflection To Load Log Implementations

However, there is actually no reason why LogFactory requires symbolic references to particular Log implementations. Reflection can be used to load these from an appropriate classloader without unacceptable performance degradation. This is the strategy adopted by JCL.

JCL uses the context classloader to load the Log implementation.