Linaro Connect Bangkok

There is a lack of unit tests available for the scheduler, energy aware scheduling, and CPU frequency management. In this session a recent effort to expand the available tests will be described and discussed.

It can be a rather complicated task to deploy optimized power management (PM) support in a driver in Linux. There are several PM frameworks and corresponding function callbacks available per device, which the driver developer needs detailed knowledge about. Particularly, when the goal is to reach the best energy efficient behavior.

Additionally, ARM SoCs in general, have quite sophisticated and fine grained methods to put parts of a silicon into a low power state, as to avoid wasting power when there are no active users of these parts. In Linux these parts are typically modeled as so called, PM domains.

During the session, we dive into some of the relevant PM frameworks for dealing with idle and explains the concepts behind them. We look into how to deploy support for system wide low power states, such as suspend to ram, suspend to idle and suspend to disk. We look at it, both from the PM domain and the driver point of view.

Moreover, to deploy fine grained PM support, the session gives some best practices of how to use runtime PM and the generic PM domain frameworks, as well as looks into how to implement support for called wakeup interrupts.

CPU Idle governor and irq prediction

Enable deepest suspend power state on S2I

Most SoC designs today implement a central entity that orchestrates System and Power Management controls. Such an entity may often be referred to as the System Control Processor (SCP), which is often a microcontroller. We use the term SCP to refer to the concept of such an entity throughout the rest of this abstract. The communication interface to the SCP is via transport channels.

The ARM System Control & Management Interface (SCMI) specification v1.0 was designed to provide an industry standard transport independent interface to the SCP to enable system and power management. SCMI enables transport channels to ferry SCMI commands from an agent sitting on a Processing Entity (like an Application Processor) to the SCP. The SCP then decodes and implements the actions described by the SCMI commands. SCMIv1.0 is an already published specification supporting power, performance, clocks, sensors and system state management.

Our presentation will talk about the vision of SCMI to enable a truly scalable Power Management software stack that can work across all compliant devices with minimal modification, with device specific controls resident in firmware. It will talk about evolution of the ARM SCMI Specification and how the next revision of the specification will enable next-generation system designs, especially touching a few key areas as outlined below.

There is a growing requirement in certain segments of the industry to deploy virtualized systems, especially in the automotive domain. We will explain how SCMI can enable Power Management virtualization in such systems and design considerations thereof. We will also introduce the concept of Reset Domains and their Management via SCMI. Currently most Operating Systems manage Device Power, Clock and Performance via various frameworks in order to control a device. We envision a unified way of managing a device via a Device-centric model of Power Management. We will explain how SCMI can enable such a model and simplify current Power Management stacks.

Our presentation will be of interest to people and organizations who are looking to enable scalable Power Management stacks for future devices where multiple Operating Systems may run concurrently in the same SoC, either physically on different Processing Entities, or via Virtualization.

Bus scaling enhancements

Non upstreamable part of EAS

Scheduler load balance rework

This training will present the details of some parts of the scheduler like the task placement during wake up path, the cgroup in the scheduler or how CPU compute capacity is used to balance tasks on the system.