Hi Laurent, On 25-Nov-24 1:49 PM, Laurent Pinchart wrote: > On Mon, Nov 25, 2024 at 01:25:41PM +0100, Hans de Goede wrote: <snip> >> I see 2 ways of doing that: >> >> 1. Use pm_runtime_set_autosuspend_delay() with a delay of say 1 second >> and then on set_ctrl do a pm_runtime_get_sync() + >> pm_runtime_put_autosuspend() giving the camera 1 second to finish >> applying the async ctrl (which might not be enough for e.g homing) + >> also avoid doing suspend + resume all the time if multiple ctrls are send >> >> 2. Instead of immediately powering on the camera on /dev/video# open >> track per fh if the camera has been powered on and then on the first >> set-ctrl, or the first other IOCTL like try/set-fmt which requires >> the camera to be powered on power it up and then keep it on until >> the fh is closed, since apps hitting these paths are likely to do >> more stuff which requires the camera to be powered on. > > A mode of operation where a userspace action causes a state change and > the only way to change back to the previous state is to close the device > often leads to problems. I'd rather not do this unless we have to > completely rule out all other options. But we already have that today. We already do the usb_autopm_get_interface() as soon as /dev/video# gets opened and the only way to undo it is to close /dev/video#. What I'm suggesting is to no longer do the usb_autopm_get_interface() on all opens, but only on some. Where "some" are the ones where we come to the conclusion that we actually need to power-up the USB-bus / interface because we want to talk to the device. IOW delay the usb_autopm_get_interface() until the first action which actually requires it. This should be a very minimal change from the pov of USB interactions with the actual device, so a small change of regressions while at least not powering on the device during udev discovery. I guess one could argue that the cases where this is a win are so small that this is not worth it. Regards, Hans
