On Mon, Oct 03, 2011 at 12:57:10PM +0200, Guennadi Liakhovetski wrote: > Hi all Hi Guennadi, Thanks for a thoughtful writing on subdev PM! > (The original .s_power() author added to cc;-)) > > Here comes one more Request for Discussion from me. > > Short: on what events, at which level and how shall subdevice PM be > envoked? > > Subdevices can have varying and arbitrarily complex Power Management > methods. On-SoC subdevices would typically be powered on and off by > writing to some system registers. External subdevices (sensors etc.) can > be powered on or off by something as simple as a GPIO, or can use several > power regulators, supplying power to different device circuits. This > means, a part of this knowledge belongs directly to the driver, while > another part of it comes from platform data. The driver itself knows, > whether it can control device's power, using internal capabilities, or it > has to request a certain number of regulators. In the latter case, > perhaps, it would be sane to assume, that if a certain regulator is not > available, then the respective voltage is supplied by the system > statically. > > When to invoke? Subdeices can be used in two cases: for configuration and > for data processing (streaming). For configuration the driver can choose > one of two approaches: (1) cache all configuration requests and only > execute them on STREAMON. Advantages: (a) the device can be kept off all > the time during configuration, (b) the order is unimportant: the driver > only stores values and applies them in the "correct" order. Disadvantages: > (a) if the result of any such operation cannot be fully predicted by the > driver, it cannot be reported to the user immediately after the operation > execution but only at the STREAMON time (does anyone know any such > "volatile" operations?), (b) the order is lost (is this important?). (2) > execute all operations immediately. Advantages and disadvantages: just > invert those from (1) above. > > So far individual drivers decide themselves which way to go. This way only > drivers themselves know, when and what parts of the device they have to > power on and off for configuration. The only thing the bridge driver can > be sure about is, that all the involved subdevices in the pipeline have to > be powered on during streaming. But even then - maybe the driver can and > wants to power the i2c circuitry off for that time? The bridge driver can't (nor should) know about the power management requirements of random subdevs. The name of the s_power op is rather poitless in its current state. The power state of the subdev probably even never matters to the bridge --- or do we really have an example of that? In my opinion the bridge driver should instead tell the bridge drivers what they can expect to hear from the bridge --- for example that the bridge can issue set / get controls or fmt ops to the subdev. The subdev may or may not need to be powered for those: only the subdev driver knows. This is analogous to opening the subdev node from user space. Anything else except streaming is allowed. And streaming, which for sure requires powering on the subdev, is already nicely handled by the s_stream op. What do you think? In practice the name of s_power should change, as well as possible implementatio on subdev drivers. > All the above makes me think, that .s_power() methods are actually useless > in the "operation context." The bridge has basically no way to know, when > and which parts of the subdevice to power on or off. Subdevice > configuration is anyway always performed, using the driver, and for > streaming all participating subdevices just have to be informed about > streaming begin and end. > > The only pure PM activity, that subdevice drivers have to be informed > about are suspends and resumes. Normal bus PM callbacks are not always > usable in our case. E.g., you cannot use i2c PM, because i2c can well be > resumed before the bridge and then camera sensors typically still cannot > be accessed over i2c. Do you have a bridge that provides a clock to subdevs? The clock should be modelled in the clock framework --- yes, I guess there's still a way to go before that,s universally possible. > Therefore I propose to either deprecate (and later remove) .s_power() and > add .suspend() and .resume() instead or repurpose .s_power() to be _only_ > used for system-wide suspending and resuming. Even for runtime PM the > subdevice driver has all the chances to decide itself when and how to save > power, so, again, there is no need to be called from outside. -- Sakari Ailus e-mail: sakari.ailus@xxxxxx jabber/XMPP/Gmail: sailus@xxxxxxxxxxxxxx -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
