Hi Jacopo, On Tue, Sep 10, 2024 at 11:42:06AM +0200, Jacopo Mondi wrote: > > > > > > > Ah, I missed that one. > > > > > > > > > > > > > > I don't think it fixes the issue I mentioned. If we have PM enabled, and the > > > > > > > parent device requires powering up for the child device (BE) to be > > > > > > > accessible, the driver will crash when calling pispbe_hw_init(). I think you > > > > > > > should call pm_runtime_set_active() before calling pispbe_runtime_resume(). > > > > > > > > > > > > As discussed, this is not a problem currently for BE, but indeed you > > > > > > have a point. > > > > > > I admit the runtime_pm intrinsics are obscure to me, but Laurent just > > > made me notice something. > > > > > > Consider the following scenario > > > > > > *) Kernel compiled with CONFIG_PM > > > *) i2c sensor driver that supports both CONFIG_PM and !CONFIG_PM by: > > > *) Manually power up the sensor during probe > > > *) Call pm_runtime_enable() and pm_runtime_set_active() at the end > > > of the probe routine after having accessed the chip over i2c > > > (like most, if not all the i2c drivers in mainline do including > > > ccs) > > > > > > All these drivers work, and during the probe routine before accessing > > > the HW, they don't need to power up the parent i2c controller. > > > > This isn't done explicitly by the I²C drivers, indeed but... > > > > > > > > Might it be that during probe() the parent is guaranteed to be enabled ? > > > > ...yes. > > > > Unrelated, but I am wrong the driver core calls pm_request_idle() on > the just probed device ? Does this mean drivers doesn't have to do > that by themseleves ? (it's not a big deal, as request_idle() doesn't > change the usage counter) Seems like that. This could be omitted from drivers then. > > > > > > > I add a look in the driver-core and pm Documentation/ but found > > > nothing. > > > > > > A quick stroll in driver/base/ got me to __device_attach() and it > > > seems parents are powered up before attaching a driver to a device > > > (which in my understanding should be what ends up calling probe()). > > > Clearly I've no real understanding of what I'm talking about when it > > > comes to driver-core, so take this with a grain of salt. > > > > This only works with runtime PM enabled. > > > > It's the CONFIG_PM case that was worrying for Tomi > > > > > > > > > > > > > > > > Sad note: most of all the occurrences of "grep set_active" in > > > > > > drivers/media/platform/ show that set_active() is used as I've done in > > > > > > my patch > > > > > > > > > > > > > However, you said above that "supporting !CONFIG_PM is not that much work". > > > > > > > Maybe not. But how much work is it to get it right (for both PM and !PM), > > > > > > > and make sure it stays right? =). > > > > > > > > > > > > > > Just my opinion, but if there are zero use cases for the !PM, I would just > > > > > > > go with "depends on PM" to keep the driver simpler, less bug-prone, and > > > > > > > easier to maintain. > > > > > > > > > > I'm fine with that, and for platform drivers, that's my preferred > > > > > option. Sakari ? > > > > > > > > I'm concerned with your (?) recent finding that many architectures don't > > > > have support for CONFIG_PM. In this case the device is very unlikely to be > > > > used outside ARM(64) so I guess it's fine. > > > > > > > > > > Also, this IP is RPi specific, and the !CONFIG_PM case is not used or > > > tested on Pi. > > > > > > However, I think this current patch is correct (assuming the above > > > reasoning on i2c sensor drivers is correct) and doesn't require > > > CONFIG_PM, so I would be tempted to keep this version. > > > > I understand the current patch does depend on CONFIG_PM: it requires > > runtime PM to be operational to start the clock, for instance. > > > > Where do you see that ? > > This version still calls pispbe_runtime_resume() to power up the IP, > it doesn't go through runtime_pm: > https://patchwork.linuxtv.org/project/linux-media/patch/20240904095836.344833-5-jacopo.mondi@xxxxxxxxxxxxxxxx/ cfe_runtime_resume() is only called via runtime PM. > > Thanks! > > > > > > > > > > > > > > > I don't see a use case for !PM and we confirmed with RPi they don't > > > > > > need to support it. During the review of a previous version of the BE > > > > > > driver iirc I've been asked to support !PM but I'm not sure I recall > > > > > > the reasons. > > > > > > > > > > I hope it wasn't me :-) > > > > > > > > Me neither. Although it'd be nice: CONFIG_PM isn't a hardware specific > > > > option as such. As one part of the kernel requires !CONFIG_PM and another > > > > CONFIG_PM, we can expect problems, at least in principle. > > > > > > > > Ideally all architectures would support it so CONFIG_PM could be removed > > > > and we could say the problem has been solved. :-) -- Kind regards, Sakari Ailus
