Re: [PATCH v2] usb:gadget:uvc Do not use worker thread to pump usb requests

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Fri, Oct 27, 2023 at 03:39:44PM +0200, Michael Grzeschik wrote:
> On Fri, Oct 27, 2023 at 02:47:52PM +0300, Laurent Pinchart wrote:
> > On Fri, Oct 27, 2023 at 01:10:21PM +0200, Michael Grzeschik wrote:
> > > On Fri, Oct 27, 2023 at 10:51:17AM +0300, Laurent Pinchart wrote:
> > > > On Thu, Oct 26, 2023 at 09:56:35PM +0000, Jayant Chowdhary wrote:
> > > >> This patch is based on top of
> > > >> https://lore.kernel.org/linux-usb/20230930184821.310143-1-arakesh@xxxxxxxxxx/T/#t:
> > > >>
> > > >> When we use an async work queue to perform the function of pumping
> > > >> usb requests to the usb controller, it is possible that thread scheduling
> > > >> affects at what cadence we're able to pump requests. This could mean usb
> > > >> requests miss their uframes - resulting in video stream flickers on the host
> > > >> device.
> > > >>
> > > >> In this patch, we move the pumping of usb requests to
> > > >> 1) uvcg_video_complete() complete handler for both isoc + bulk
> > > >>    endpoints. We still send 0 length requests when there is no uvc buffer
> > > >>    available to encode.
> > > >
> > > > This means you will end up copying large amounts of data in interrupt
> > > > context. The work queue was there to avoid exactly that, as it will
> > > > introduce delays that can affect other parts of the system. I think this
> > > > is a problem.
> > > 
> > > Regarding Thin's argument about possible scheduling latency that is already
> > > introducing real errors, this seemed like a good solution.
> > > 
> > > But sure, this potential latency introduced in the interrupt context can
> > > trigger other side effects.
> > > 
> > > However I think we need some compromise since both arguments are very valid.
> > 
> > Agreed.
> > 
> > > Any ideas, how to solve this?
> > 
> > I'm afraid not.
> 
> We discussed this and came to the conclusion that we could make use of
> kthread_create and sched_setattr with an attr->sched_policy = SCHED_DEADLINE
> here instead of the workqueue. This way we would ensure that the worker
> would be triggered with hard definitions.
> 
> Since the SG case is not that heavy on the completion handler, we could
> also make this kthread conditionaly to the memcpy case.

If you don't mind a naive suggestion from someone who knows nothing 
about the driver...

An attractive possibility is to have the work queue (or kthread) do the 
time-consuming copying, but leave the submission up to the completion 
handler.  If the data isn't ready (or there's no data to send) when the 
handler runs, then queue a 0-length request.

That will give you the best of both worlds: low latency while in 
interrupt context and a steady, constant flow of USB transfers at all 
times.  The question of how to schedule the work queue or kthread is a 
separate matter, not directly relevant to this design decision.

Alan Stern




[Index of Archives]     [Linux Media]     [Linux Input]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [Old Linux USB Devel Archive]

  Powered by Linux