Re: uvc gadget performance issues with skip interrupt impl

[Greg Kroah-Hartman]

On Tue, Oct 25, 2022 at 01:34:01AM -0500, Jeff Vanhoof wrote:
> Hi,
> 
> During the queuing up of requests from the UVC Gadget Driver to DWC3 for one
> frame, if a missed isoc event occurs then it is possible for the next
> consecutive frame(s) to also see missed isoc related errors as a result,
> presenting to the user as a large video stall.
> 
> This issue appears to have come in with the skip interrupt implementation in
> the UVC Gadget Driver:
> 
> usb: gadget: uvc: decrease the interrupt load to a quarter
> https://lore.kernel.org/r/20210628155311.16762-6-m.grzeschik@xxxxxxxxxxxxxx
> 
> Below is an example flow of how the issue can occur (and why).
> 
> For example (ISOC use case):
> 1) DWC3 driver has 4 requests queued up from the UVC Gadget Driver.
> 
> 2) First request has IOC bit set due to no_interrupt=0 also being set, and IMI
> bit is set to detect missed ISOC.
> 
> 3) Requests 2,3,4 do not have IOC bit set due to no_interrupt=1 being set for
> them. (Note: Whether or not the IMI bit is set for these requests does not
> matter, issue can still crop up as there is no guarantee that request 2,3,4
> will see a missed isoc event)
> 
> 4) First request gets a missed isoc event and DWC3 returns the req and error to
> UVC Gadget Driver.
> 
> 5) UVC Gadget Driver, in uvc_video_complete, proceeds to cancel the queue by
> calling uvcg_queue_cancel.
> 
> 6) UVC Gadget Driver stops sending additional requests for the current frame.
> 
> 7) DWC3 will still have requests 2,3,4 queued up and sitting in its
> started_list as these requests are not given back to the UVC gadget driver
> because they each have no_interrupt=1 set, and the DWC3 driver will not have
> any additional interrupts triggered for them as a result.
> 
> 8) Approximately 30-100ms later a new frame enters the UVC Gadget Driver (from
> V4L2), and it proceeds to send additional requests to the DWC3 driver.
> 
> 9) Because requests 2,3,4 are still sitting in the started_list of the dwc3
> driver, the driver does not stop and restart the transmission that normally
> helps it recover from the missed isoc situation (this usually happens in
> between frames).
> 
> 10) Some of the requests from the new frame will have no_interrupt=0 set, but
> these requests will be considered missed/late by the DWC3 controller.
> 
> 11) Because these new requests have the IOC bit set (and possibly IMI),
> interrupts will be triggered causing the DWC3 Driver to return the req and
> error to the UVC Gadget Driver.
> 
> 12) And if the last set of requests sent by the UVC Gadget Driver have
> "no_interrupt=1" set, then DWC3 may not interrupt further until new requests
> come in, and the cycle of frame drops/errors will continue.
> 
> I have briefly mentioned this issue in another conversation with Thinh. At the
> time he mentioned that 3 things could possibly be done to help resolve this
> issue:
> 
> 1) The UVC Gadget Driver should ensure that the last requests queued to DWC3
> must always have "no_interrupt=0" set.
> 
> 2) DWC3 can detect stale requests, stop the transmission and give back the
> requests to the UVC Gadget Driver, and restart the transmission for the new set
> of requests.
> 
> 3) Set "no_interrupt=0" for each request.
>  
> I have tested out various implementations for all 3 possibilities and they each
> seem to work ok. Note that these test implementations are not ready for prime
> time, but served as a way to prove that potential changes in these areas could
> help to resolve this issue.
> 
> I believe that a change for the UVC Gadget Driver should be made, but it also
> makes sense for the DWC3 driver to also attempt to recover from this situation
> if possible.
> 
> Does anyone have an opinion on the best way to proceed?

Please see this set of patches and the discussion around them:
	https://lore.kernel.org/r/20221018215044.765044-1-w36195@xxxxxxxxxxxx

Some of them are already queued up in my tree and in linux-next, can you
try that?  There are others for the dwc3 driver on the mailing list as
well, testing those would be wonderful if you could do that.

thanks,

greg k-h