Re: uvc gadget: Making upper bound of number of usb requests allocated configurable through configfs

[Laurent Pinchart <laurent.pinchart@ideasonboard.com>]

On Fri, Oct 20, 2023 at 11:30:52PM +0000, Thinh Nguyen wrote:
> Sorry for the delay response.
> 
> On Sun, Oct 15, 2023, Jayant Chowdhary wrote:
> > On 10/12/23 11:50, Thinh Nguyen wrote:
> > > The frequency of the request submission should not depend on the
> > > video_pump() work thread since it can vary. The frequency of request
> > > submission should match with the request completion. We know that
> > > request completion rate should be fixed (1 uframe/request + when you
> > > don't set no_interrupt). Base on this you can do your calculation on how
> > > often you should set no_interrupt and how many requests you must submit.
> > > You don't have to wait for the video_pump() to submit 0-length requests.
> > >
> > > The only variable here is the completion handler delay or system
> > > latency, which should not be much and should be within your calculation.
> > 
> > 
> > Thanks for the suggestion. It indeed makes sense that we do not completely depend on
> > video_pump() for sending 0 length requests. I was concerned about
> > synchronization needed when we send requests to the dwc3 controller from
> > different threads. I see that the dwc3 controller code does internally serialize
> > queueing requests, can we expect this from other controllers as well ? 
> 
> While it's not explicitly documented, when the gadget driver uses
> usb_ep_queue(), the order in which the gadget recieves the request
> should be maintained and serialized. Because the order the transfer go
> out for the same endpoint can be critical, breaking this will cause
> issue.

That's right, but if usp_ep_queue() is called from multiple contexts,
there's no guarantee it can provide when it comes to the ordering. The
caller has to handle it.

> > This brings me to another question for Michael - I see
> > that we introduced a worker thread for pumping  usb requests to the usb endpoint
> > in https://urldefense.com/v3/__https://lore.kernel.org/all/20200427151614.10868-1-m.grzeschik@pengutronix.de/__;!!A4F2R9G_pg!aAnzCopbTbZtUxBK6a6r6_QzV-b2Z2J5o5esPaartZ2jogKijmhqj6OyiKDg-BPhxq8pJHR3HS1Hf8z6YnqfWTon$ 
> > (I see multiple email addresses, so apologies if I used the incorrect one).
> > 
> > Did we introduce the worker thread to solve some specific deadlock scenarios ?
> > Or was it a general mitigation against racy usb request submission from v4l2 buffer
> > queuing, stream enable and the video complete handler firing ?
> > 
> > I was chatting with Avi about this, what if we submit requests to the endpoint
> > only at two points in the streaming lifecycle - 
> > 1) The whole 64 (or however many usb requests are allocated) when
> >    uvcg_video_enable() is called - with 0 length usb_requests.
> > 2) In the video complete handler - if a video buffer is available, we encode it
> >    and submit it to the endpoint. If not, we send a 0 length request.
> > 
> > This way we're really maintaining back pressure and sending requests as soon
> > as we can to the dwc3 controller. Encoding is mostly memcpys from what I see
> > so hopefully not too heavy on the interrupt handler. I will work on prototyping
> > this meanwhile.
> 
> That sounds good to me. I believe Michael already provided some test
> patches and you've already done some preliminary tests for that right?

-- 
Regards,

Laurent Pinchart