Hi Ricardo, On 25-Nov-24 2:39 PM, Ricardo Ribalda wrote: > On Mon, 25 Nov 2024 at 13:25, Hans de Goede <hdegoede@xxxxxxxxxx> wrote: >> >> Hi Ricardo, >> >> On 9-Nov-24 5:29 PM, Ricardo Ribalda wrote: >> >> <snip> >> >>>> I have been discussing UVC power-management with Laurent, also >>>> related to power-consumption issues caused by libcamera's pipeline >>>> handler holding open the /dev/video# node as long as the camera >>>> manager object exists. >> >> <snip> >> >>>> Here is what I have in mind for this: >>>> >>>> 1. Assume that the results of trying a specific fmt do not change over time. >>>> >>>> 2. Only allow userspace to request fmts which match one of the enum-fmts -> >>>> enum-frame-sizes -> enum-frame-rates tripplet results >>>> (constrain what userspace requests to these) >>>> >>>> 3. Run the equivalent of tryfmt on all possible combinations (so the usaul >>>> 3 levels nested loop for this) on probe() and cache the results >>>> >>>> 4. Make try_fmt / set_fmt not poweron the device but instead constrain >>>> the requested fmt to one from our cached fmts >>>> >>>> 5. On stream-on do the actual power-on + set-fmt + verify that we get >>>> what we expect based on the cache, and otherwise return -EIO. >>> >>> Can we start powering up the device during try/set fmt and then >>> implement the format caching as an improvement? >> >> Yes, actually looking at how complex this is when e.g. also taking >> controls into account I think that taking small steps is a good idea. >> >> I have lately mostly been working on sensor drivers where delaying >> applying format settings + all controls to stream-on is normal. >> >> So that is the mental model I'm applying to uvc here, but that might >> not be entirely applicable. >> >>> Laurent mentioned that some cameras missbehave if a lot of controls >>> are set during probing. I hope that this approach does not trigger >>> those, and if it does it would be easier to revert if we do the work >>> in two steps. >> >> Ack, taking small steps sounds like a good plan. >> >> <snip> >> >>>> This should also make camera enumeration faster for apps, since >>>> most apps / frameworks do the whole 3 levels nested loop for this >>>> on startup, for which atm we go out to the hw, which now instead >>>> will come from the fmts cache and thus will be much much faster, >>>> so this should lead to a noticeable speedup for apps accessing UVC >>>> cameras which would be another nice win. >>>> >>>> Downside is that the initial probe will take longer see we do >>>> all the tryfmt-s there now. But I think that taking a bit longer >>>> to probe while the machine is booting should not be an issue. >>> >>> How do you pretend to handle the controls? Do you plan to power-up the >>> device during s_ctrl() or set them only during streamon()? >>> If we power-up the device during s_ctrl we need to take care of the >>> asynchronous controls (typically pan/tilt/zoom), The device must be >>> powered until the control finishes, and the device might never reply >>> control_done if the firmware is not properly implemented. >>> If we set the controls only during streamon, we will break some >>> usecases. There are some video conferencing equipment that do homing >>> during streamoff. That will be a serious API breakage. >> >> How to handle controls is a good idea. >> >> Based on my sensor experience my initial idea was to just cache them >> all. Basically make set_ctrl succeed but do not actually do anyhing >> when the camera is not already powered on and then on stream-on call >> __v4l2_ctrl_handler_setup() to get all current values applied. >> >> But as you indicate that will likely not work well with async controls, >> although we already have this issue when using v4l2-ctl from the cmdline >> on such a control and that seems to work fine. > > ----- >> Just because we allow >> the USB connection to sleep, does not mean that the camera cannot finish >> doing applying the async control. >> > Not sure what you mean with this sentence. Could you explain it > differently? Sorry > >> But I can see how some cameras might not like this and having 2 different >> paths for different controls also is undesirable. >> >> Combine that with what Laurent said about devices not liking it when >> you set too much controls in a short time and I do think we need to >> immediately apply ctrls. >> >> 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 > > What about 1.5: > > during s_ctrl(): > usb_autopm_get_interface() > if the control is UVC_CTRL_FLAG_ASYNCHRONOUS. > usb_autopm_get_interface() > set the actual control in the hardware > usb_autopm_put_interface() > > during uvc_ctrl_status_event(): > usb_autopm_put_interface() How do we match this to the usb_autopm_get_interface() call ? At a minimum we would need some counter to track pending (not acked through status interrupt urb) async control requests and only do the put() if that counter >= 1 (and then decrease the counter). We don't want to do unbalanced puts here in case of buggy cameras sending unexpected / too many ctrl status events. > during close(): > send all the missing usb_autopm_put_interface() Except for my one remark this is an interesting proposal. Maybe also do a dev_warn() if there are missing usb_autopm_put_interface() calls pending on close() ? > This way: > - we do not have an artificial delay that might not work for all the use cases > - cameras with noncompliant async controls will have the same PM > behaviour as now (will be powered on until close() ) > > We do the same with the rest of the actions that require hardware access, like: > https://lore.kernel.org/linux-media/20220920-resend-powersave-v5-2-692e6df6c1e2@xxxxxxxxxxxx/ > > This way: > - Apps that do not need to access the hardware, do not wake it up, and > we do not break usecases. > > Next steps will be: > - cache the formats > - move the actual set_ctrl to streamon... but if we can do that I > would argue than we can move completely to the control framework. Right I had forgotten that the UVC driver does not use the control framework. I think moving to that would be a prerequisite for moving the set_ctrl to stream_on. Regards, Hans
