[libcamera-devel] [RFC v1 5/7] py: Add 'nonblocking' argument to get_ready_requests()

[David Plowman]

Hi everyone

Sorry for not taking part in this discussion rather more, and thanks
to those who have! I think I was asked further back if I had an
opinion, so let me try and explain how I see things.

Actually I don't have any particularly strong opinions, other than
that I want things to work and to be easy to use. And if they stopped
changing, that would be nice too!!

I was actually OK with the previous version where get_ready_requests
was non-blocking, it's only the more recent version that causes me
some trouble because I can't simply flush out any lurking requests
after stopping the camera - because there might not be any and I will
simply lock up the system.

It's perhaps worth noting that, like our C++ apps, I recycle requests
back to libcamera just as soon as is humanly possible, as that's the
best way to avoid frame drops. The number of request objects that I
make is typically from 1 to about 8 (inclusive).

I'm happy to go back to the non-blocking version, or I'm happy to have
a blocking version where I'm guaranteed not to get any more requests
out once the camera has stopped. Once the camera is stopped I just
discard all my requests and don't want to see them again! I suppose if
they came out later but were marked as "cancelled" that would be
workable, though it feels less tidy.

I would dislike the approach where I have to keep track of which
requests I've sent and which ones haven't yet come back, so that I
know whether there are lurking requests that I have to wait for. Not
that it's difficult, it's just annoying, and I expect it would be a
nuisance to very many Python users. It might also be slow if there are
quite a few requests still to be fulfilled (depending on whether they
can be "cancelled" quickly).

I guess there are other solutions where maybe I could test if the file
descriptor is readable and so on, but again, why would we make it
complicated like this?

I did notice the comment earlier that the Python bindings are far from
stable. I certainly agree that where things are unfinished or untidy,
those things need to be sorted out, but in other areas we basically
have what we need for all our functionality, and so would be happy to
see less churn. I don't know if that's possible in the short term...

So maybe I did have some opinions after all!

Thanks everyone
David

On Tue, 28 Jun 2022 at 08:08, Tomi Valkeinen
<tomi.valkeinen at ideasonboard.com> wrote:
>
> On 27/06/2022 23:42, Laurent Pinchart wrote:
> > Hi Tomi,
> >
> > On Mon, Jun 27, 2022 at 03:54:11PM +0300, Tomi Valkeinen wrote:
> >> On 27/06/2022 15:25, Laurent Pinchart wrote:
> >>> On Mon, Jun 27, 2022 at 01:51:36PM +0300, Tomi Valkeinen wrote:
> >>>> On 27/06/2022 13:16, Laurent Pinchart wrote:
> >>>>> On Mon, Jun 27, 2022 at 12:49:45PM +0300, Tomi Valkeinen wrote:
> >>>>>> On 24/06/2022 16:53, Laurent Pinchart wrote:
> >>>>>>> On Fri, Jun 24, 2022 at 11:26:18AM +0100, David Plowman wrote:
> >>>>>>>> Hi everyone
> >>>>>>>>
> >>>>>>>> Just to add some background to this one...
> >>>>>>>>
> >>>>>>>> This has been causing me a little trouble in Picamera2 lately.
> >>>>>>>> get_ready_requests used to be non-blocking so I would always call it
> >>>>>>>> after stopping the camera to clear out any "lurking" requests. Since
> >>>>>>>> it became blocking I've had to stop that but it does mean that a stray
> >>>>>>>> request can be read out at an awkward time. This can lead to us
> >>>>>>>> getting the "wrong" image (and probably falling over when it's the
> >>>>>>>> wrong size or something), or even trying to queue it back to libcamera
> >>>>>>>> (while the camera is still stopped).
> >>>>>>>>
> >>>>>>>> Anyway, either solution works for me. Either I can flush out those
> >>>>>>>> requests after calling stop(), which is what I used to do. Or they
> >>>>>>>> could disappear "spontaneously". Both work for me!!
> >>>>>>>
> >>>>>>> I'd like to make get_ready_requests() non-blocking unconditionally,
> >>>>>>> could that be done ?
> >>>>>>
> >>>>>> I think we should have a blocking version too.
> >>>>>
> >>>>> Why is that ?
> >>>>
> >>>> It's the easiest and the most natural way to write a small script to get
> >>>> something captured. Maybe we'll do it with a separate function, but
> >>>> somehow we need to offer a way to capture without dealing with Selectors
> >>>> or such.
> >>>
> >>> I'd prefer, if possible, to offer higher-level features on top of the
> >>> core bindings. By excluding them from the core bindings, I think it will
> >>> leave room for people to experiment with higher-level camera APIs, be it
> >>> in the Raspberry Pi camera Python code, or in other implementations.
> >>
> >> That is a valid point, but I think mine is too =). Especially with a
> >> feature like this, which doesn't hide or take anything away. I don't see
> >> what a blocking version would take away from a higher level API.
> >>
> >> I'll do some experimenting, perhaps it's so trivial to use a
> >> non-blocking version with Selector or something else that we can just
> >> drop the blocking version.
> >
> > If it's not trivial then maybe it means the core bindings are not good
> > enough :-) Implementing this purely in Python would be a good test.
>
> Oh, it's trivial in Python in the sense that it's a few lines of code.
> But it's not trivial in the sense that you could just write
>
> while True:
>         # This waits until there are requests
>          reqs = cm.get_ready_requests()
>          for req in reqs:
>                 do something
>
> We can easily write a helper function on top of non-blocking API to
> achieve the above.
>
> >> But I do think it's important to easily provide that feature, even with
> >> the core bindings. Unless we want to target the core bindings only as a
> >> base for higher level libraries, which I don't think is a good idea.
> >
> > It seems to be time to start discussing this. So far, I would prefer
> > having the core bindings matching the C++ API as closely as possible,
> > without additional features, and having convenience helpers developed on
> > top. We could develop our own convenience helpers shipped with libcamera
> > of course.
>
> I can't make my mind on this. On one hand, what you suggest is a clean
> approach. On the other, if we have a core bindings module and a
> convenience module on top, it brings up some questions:
>
> - Module naming. libcamera-core and libcamera? libcamera and
> libcamera-for-dummies?
>
> - Is the convenience module just a sugar topping, i.e. helper
> functions/classes here and there, or is it something that fully wraps
> the core (somewhat like pilibcamera2)?
>
> - If it's just some helpers, what is the benefit of using the core module?
>
> - If it fully wraps the core, most likely it somewhat dummies down the
> libcamera features. If so, there are always users for the core module.
> Those users would most likely want to use simple helpers like blocking
> wait or mmapped fb.
>
> - Perhaps a third module option would be something in between: a module
> that "pythonizes" the classes, while still exposing everything. But are
> the convenience features part of that, or yet another module on top?
> Such a module also brings up the annoyance that you would access many
> things via the higher level module (say, pylibcamera.Camera), but
> probably many things would be just exposed from libcamera module if we
> don't do a full wrap (say, camera.set_format(libcamera.formats.RGB123).
>
> So... I think it's clear that a fully wrapping module is obviously a new
> module on top of the core bindings. But the simple helpers and possibly
> the pythonization could be part of the core bindings, or on top.
>
> It's clear that features like mmapped fb class and blocking wait are
> features that are needed by the users, so I think it's not a question of
> should we have those, but where should they be.
>
> Also, not that even if the helpers would be part of the core bindings,
> we could have internal split there: the bindings made in C++ would be
> the core bindings, and in the same module we could have convenience
> features written in Python (similar to the MappedFrameBuffer.py we have
> now). However, while that sounds nice, I don't know if there's any real
> logic to it. If something is easier/better written in the C++ bindings,
> I don't see a point in forcing that feature to be written in Python instead.
>
> > Anyone else wants to chime in ?
> >
> >>>>>>>>>> Perhaps the camera.start() should go and discard any old Requests
> >>>>>>>>>> related to that camera. For the exit issue I don't see any automatic
> >>>>>>>>>> solution.
> >>>>>>>>>
> >>>>>>>>> At the end of camera.stop() we should validate and ensure that all
> >>>>>>>>> Requests are released to the application. We should hold no internal
> >>>>>>>>> requests when camera.stop() completes.
> >>>>>>>>>
> >>>>>>>>> I.e. ... if we have things to discard at camera.start() - that's a bug I
> >>>>>>>>> believe.
> >>>>>>>>>
> >>>>>>>>> ahhh but here perhaps the issue is that the python code is the one that
> >>>>>>>>> has to 'retrieve' those, while in C++ they are returned via a signal?
> >>>>>>>>>
> >>>>>>>>> Is there any harm in discarding the requests at the end of camera.stop()
> >>>>>>>>> such that there is simply 'nothing left to process' after? Or does the
> >>>>>>>>> python application have to end up owning the requests to correctly
> >>>>>>>>> release them?
> >>>>>>>
> >>>>>>> Sounds like an idea to explore. What's the drawback of clearing in
> >>>>>>> stop() ?
> >>>>>>
> >>>>>> Losing events that you expected to get, I think.
> >>>>>>
> >>>>>> I'm talking here about the event handling after adding the new event
> >>>>>> handling, as I think it's more relevant than figuring out how to do
> >>>>>> things with just the single event.
> >>>>>>
> >>>>>> We could dispatch the events (All events? Or just events related to
> >>>>>> Requests for that camera?) automatically in stop(), but that would break
> >>>>>> the backward compatibility.
> >>>>>
> >>>>> I don't recall if I've mentioned it in the review of another patch in
> >>>>> the series, but I've been thinking about dispatching events at stop
> >>>>> time, yes. This is how libcamera operates for buffer and request
> >>>>> completion events, doing the same in Python would make the behaviour
> >>>>> consistent. Backward compatibility isn't a concern, the Python bindings
> >>>>> are experimental, we shouldn't let that block the design of a good API.
> >>>>
> >>>> With C++, it sounds fine as the signals are fired behind the scenes. In
> >>>> the Python bindings there's a specific call to dispatch the events.
> >>>> Implicitly dispatching the events elsewhere can be confusing.
> >>>
> >>> The explicit dispatching call is needed to work around a Python
> >>> limitation related to threads. I'm fine with that, but I'd like to keep
> >>> it as much as an internal detail as possible, thus minimizing its
> >>> explicit usage from applications. I get your point about this being
> >>> possibly confusing for users. I'd be interested in feedback from said
> >>> users :-)
> >>
> >> Me too, but this is a rather difficult question to answer, so I don't
> >> really expect to get help here =).
> >
> > Maybe David could shine some light here, as our mean user of the core
> > bindings at this point ?
> >
> >> I do like the idea of somehow forcibly making the events handled at
> >> camera.stop() time, though.
> >>
> >> Although there may be other events queued anyway. Say, camera_added.
> >> Those will stay in the event queue, holding references to the relevant
> >> objects, until the user calls dispatch_events().
> >
> > We could only dispatch buffer and request completion events in stop()
> > for the camera being stopped, that would be fine with me. I'm not sure
> > if we would gain much from such a selective dispatch though, but it
> > would certainly mimic the C++ API.
> >
> >> Then again those don't cause similar problems as getting "old" Request
> >> events after restarting a camera, so maybe they are fine.
> >>
> >> If the user wants a clean exit, he needs to either dispatch or discard
> >> those events, otherwise the cameras and camera managers will be kept
> >> alive at the app exit time.
> >
> > Maybe a cleanup function on the camera manager would do the job ? We're
> > going back to explicit start/stop then though :-)
>
> The discard_events() is essentially a cleanup function.
>
> Perhaps we can do some clever weak-ref tricks there, though... I think
> the events should not keep anything alive. Oh, but we depend on the
> event keeping the Request alive, so that doesn't work.
>
>   Tomi