Kernel Korner - AEM: a Scalable and Native Event Mechanism for Linux

In a previous article [“An Event Mechanism for Linux”, LJ, July 2003], we introduced the necessity for Linux to adopt a native and generic event mechanism in the context of telecom. Many existing solutions attempting to increase the capabilities of Linux have failed thus far. Others did not reach our level of satisfaction, because carrier-grade platforms have different levels of real-time requirements. In order to succeed, such an event mechanism must be bound tightly to the host operating system and take advantage of its capabilities in order to deliver better performance.

At the Open Systems Lab (Ericsson Research) in Montréal, Canada, we started a project in 2001 to develop a generic solution, the Asynchronous Event Mechanism (AEM). AEM allows an application to define and register callback functions for some specific events and lets the operating system execute these routines asynchronously when the events have been activated.

AEM provides an event-driven methodology of development. This is achieved through the definition of a natural user interface in which event handlers contain in their parameter lists all of the data necessary for their execution sent directly by the kernel.

AEM also is motivated by the fact that complex distributed applications based on multithreaded architectures have proven difficult to develop and port from one platform to another because of the management layer. The objective of AEM is not only to reduce the software's development time, but also to simplify source code generation in order to increase portability between different platforms and to increase the software's life cycle.

The biggest challenge of this project was designing and developing a flexible framework such that adding or updating a running system with new event-handling implementations is possible. The constraint was to be able to carry out system maintenance without rebooting the system. The modular architecture of AEM offers such capabilities.

AEM is a complementary solution to other existing notification mechanisms. One of its great benefits is the possibility to mix event-driven code and other sequential codes.

AEM is composed of one core module and a set of loadable kernel modules providing some specific event service to applications, including soft timers and asynchronous socket interfaces for TCP/IP (Figure 1). This flexible architecture permits AEM capabilities to be extended at will.

There is no restriction on what a module can implement, because each exports a range of independent pseudo-system calls to applications. In fact, this allows two different modules to make available the same functionality at the same time. Interestingly, this offers the possibility of loading a new module to implement an improved revision without breaking other applications—they continue to use the older version. This design provides the ability to load the necessary AEM modules depending on the applications' need or to upgrade modules at runtime.

Figure 1. AEM is based on one core kernel module providing the basic event functionalities and a set of independent kernel modules providing asynchronous event services to applications.

Figure 2. Architecture behind event activation and process notification in AEM. Event wait queues containing sleeping jobs are scanned, and all concerned jobs wake up at the occurrence of an event. It immediately follows the activation of the corresponding event for each related process.

The condition for such flexibility is the presence of event activation points located at strategic places in the kernel (Figure 2). Each activation point is a specific AEM queue used to activate events. In the following sections, we describe the internals of AEM in detail.

The concept of asynchrony is a major problem when the main flow of a program's execution is broken without warning in order to execute event handlers. Input requests then are handled without the knowledge of previous input states. These are delivered directly by the core kernel or the interrupt handlers and are received without presuming any kind of order. This situation constitutes a problem for some applications, including those based on TCP/IP, which rely on the transaction state to proceed.

AEM is a three-layer architecture composed of a set of pseudo-system calls for event management, a per-process event_struct performing event serialization and storing context information in order to execute user callbacks and a per-event job_struct performing event activation.