Richard J Moore - IBM

RelayFS: Transmitting data from kernel to userspace

Paper (PDF)

Linux currently has several mechanisms for relaying information about the system and applications to the user. Examples include printk and other syslog events, evlog, ltt, and oprofile. Each subsystem has its own method for relaying information from the kernel to user space and some of these mechanisms have difficulties, e.g. printks get lost. Additionally, the replication of code and maintenance is undesirable.

In this paper we describe relayfs, a unified, efficient and reliable high-speed data relay mechanism. relayfs has a simple design intended to meet the needs of current subsystems and, should other subsystems need additional services, it has the flexibility to be expanded. It allows subsystems to log data efficiently and safely using lockless technology that is designed to scale well across multiprocessor systems.

We discuss the architecture, implementation, and usage of relayfs. relayfs uses channels that allow data to be directed to a suitable buffer or buffers for the subsystems that allocated the channel. We describe the kernel API and file naming conventions, address init-time issues, and discuss performance tradeoffs available using relayfs. Finally, via example, we demonstrate how subsystems use relayfs by describing how ltt and syslog use relayfs to log their data.

Contributors: Michel Dagenais (Polytechnique Montreal), Richard Moore (IBM Linux Technical Center), Robert Wisniewski (IBM Research), Karim Yaghmour (Opersys Inc), and Thomas Zanussi (IBM Linux Technical Center).