On Tue, Apr 26, 2005 at 02:32:29PM +1000, Benjamin Herrenschmidt wrote: -->snip > > I don't like this notion of "stop" separated from power states anyway, I > think it just doesn't work in practice. Yeah, after giving it some additional thought, I think there are better ways. > > Ben. > Ok, here's a new idea. For many devices "->suspend" and "->resume" with pm_message_t is exactly what we need. However, as we support more advanced power management features, such as runtime power management, or power containers, we need something a little more specific. The exact power state must be specified among other issues. We might do something like this: Keep "->suspend" and "->resume" around unchanged. (so the states would probably remain as PMSG_FREEZE and PMSG_SUSPEND). If the driver doesn't support the more advanced PM methods just use these. They work well enough for system sleep states etc. Alternatively drivers could support a more rich power management interface via the following methods: change_state - changes a device's power state change_state(struct device * dev, pm_state_t state, struct system_state * sys_state, int reason); @dev - the device @state - the target device-specific power state @sys_state - a data structure containing information about the intended global system power state @reason - why the state must be changed (ex. RUNTIME_PM, SYSTEM_SLEEP, SYSTEM_RESUME, etc.) halt - acts somewhat like PMSG_FREEZE, stops device activity, doesn't change power state halt(struct device * dev, struct system_state * sys_state, int reason); @dev - the device @sys_state - a data structure containing information about the intended global system power state @reason - why we are halting operation (ex. RUNTIME_CHANGES (like cpufreq), SYSTEM_SLEEP, SHUTDOWN, REBOOT) contine - resumes from a "halt" continue(struct device * dev, struct system_state * sys_state, int reason); @dev - the device @sys_state - a data structure containing information about the intended global system power state @reason - why we are resuming operation (ex. RUNTIME_CHANGES (like cpufreq), SYSTEM_RESUME) When changing system state, we call "change_state" for every device with power resources. Devices that do not directly consume power or have power states will not implement "change_state" so we will call "halt" and "continue" instead. When shutting down the system, halt has the option of turning off the device, as it will see the SHUTDOWN reason. So it's a driver-knows-best approach instead of assuming everything must be turned off, or everything must just be stopped. So in theory, with cpufreq, we could stop userspace, ->halt every device (drivers won't do anything if they know it's not necessary), change the frequency, and then resume operation. We may want to create structures like pm_message_t for "change_state", "halt", and "continue". Pavel, do you have any thoughts on this? This is just a rough idea... I look forward to any comments or suggestions. Thanks, Adam
