LCA13: Combining Runtime PM and suspend/resume
4 likes3,252 views
The document discusses the integration of runtime power management (PM) with suspend/resume functionalities in device drivers, emphasizing that runtime PM allows devices to suspend on inactivity and resume as needed. It outlines the implementation challenges and solutions, such as managing power domains and ensuring devices are suspended properly to avoid being left runtime-enabled. The author encourages a better understanding of PM core protections and the proliferation of power management domains.
![Use "late" callbacks
Ensure all devices are suspended
struct dev_pm_ops {
[...]
int (*suspend_late)(struct device *dev);
int (*resume_early)(struct device *dev);
};
struct dev_pm_domain {
struct dev_pm_ops ops;
};](https://image.slidesharecdn.com/runtimepm-140321043001-phpapp02/85/LCA13-Combining-Runtime-PM-and-suspend-resume-20-320.jpg)
LCA13: Combining Runtime PM and suspend/resume
- 1. Combining Runtime PM and suspend/resume Kevin Hilman khilman@linaro.org For an Introduction to runtime PM http://people.linaro.org/~khilman/runtime_PM.html
- 2. What ? !!! Now I have to add Runtime PM to my driver? I thought suspend/resume was enough !!
- 3. System PM Runtime PM Today: Separate worlds
- 5. Runtime PM · suspend on inactivity · resume when needed
- 6. System PM · suspend when requested · resume with system
- 7. Hey! what if I could implement runtime PM, and get suspend for free?
- 8. So... how to combine them?
- 9. First, make runtime PM work. Use PM domains to manage common resources (clocks, mux, power domains, IP common...) Then, add suspend/resume... if needed.
- 10. Suspend triggers runtime PM · Force inactivity in suspend · device is now idle/inactive · runtime PM takes over
- 11. Don't implement suspend Necessary if device is used by other devices during suspend
- 12. Example: OMAP I2C driver Needed during suspend by other drivers · I2C-connected RTC · regulators on I2C-connected PMIC Solution: · runtime PM only (no suspend/resume hooks) · full (re)init per xfer; stateless Quiz 1: What if I2C driver implemented suspend and was suspended before RTC? Quiz 2: Who decides suspend ordering?
- 13. Great, so... runtime PM, plus a little "enforcement" in ->suspend() gets me both… right? Yes... Well... Almost...
- 14. Except... there are several ways for devices to be kept awake.
- 15. Tricky: async callbacks PM workqueue frozen during suspend · Asynchronous callbacks happen via PM core workqueue (pm_wq) · workqueue frozen during suspend · Result: no async callbacks Solution: use _sync() methods
- 16. Tricky: Autosuspend timeouts Driver does some work during ->suspend() · _get_sync() · _put_sync_autosuspend() · autosuspend callbacks are asynchronous · Result: device left runtime enabled Solution: PM domain magic
- 17. Tricky: PM core "protections" PM core: device_prepare() · increments usecount, decrements in _complete() · effectively disables runtime suspend (but not resume) Example: device runtime suspended: usecount = 0 · PM core prepare: _get_noresume(): usecount = 1 · driver ->suspend(): _get_sync(): usecount = 2 · driver uses HW · driver ->suspend(): _put_sync(): usecount = 1 Result: stuck with usecount > 0, runtime enabled Solution: PM domain magic
- 18. Tricky: userspace disable runtime PM can be disabled from userspace # echo on > /sys/devices/.../power/control Oops, drivers using only runtime PM will be left runtime enabled during suspend. Solution: PM domain magic
- 19. PM domain magic
- 20. Use "late" callbacks Ensure all devices are suspended struct dev_pm_ops { [...] int (*suspend_late)(struct device *dev); int (*resume_early)(struct device *dev); }; struct dev_pm_domain { struct dev_pm_ops ops; };
- 21. The magic trick Runtime suspend "enforced" for stuck devices static int my_domain_suspend_late(struct device *dev)> { if (!pm_runtime_status_suspended(dev)) { if (pm_generic_runtime_suspend(dev) == 0) { /* platform-specific PM */ } } return 0; }
- 22. Summary · Runtime PM first · use PM domains · Keep drivers generic
- 23. What next? · better understand PM core protections in device_prepare() · Help PM domains proliferate · What else?
- 24. ¿ Questions ?
- 25. 結束