Re: [RFC PATCH v2 09/22] ASoC: qcom: qdsp6: Introduce USB AFE port to q6dsp

[Wesley Cheng]

Hi Pierre,

On 1/31/2023 7:02 PM, Pierre-Louis Bossart wrote:
> On 1/31/23 20:40, Wesley Cheng wrote:
> > Hi Pierre,
> >
> > On 1/30/2023 3:59 PM, Pierre-Louis Bossart wrote:
> > > On 1/30/23 16:54, Wesley Cheng wrote:
> > > > Hi Pierre,
> > > >
> > > > On 1/26/2023 7:38 AM, Pierre-Louis Bossart wrote:
> > > > > On 1/25/23 21:14, Wesley Cheng wrote:
> > > > > > The QC ADSP is able to support USB playback endpoints, so that the
> > > > > > main
> > > > > > application processor can be placed into lower CPU power modes.  This
> > > > > > adds
> > > > > > the required AFE port configurations and port start command to
> > > > > > start an
> > > > > > audio session.
> > > > > >
> > > > > > Specifically, the QC ADSP can support all potential endpoints that are
> > > > > > exposed by the audio data interface.  This includes, feedback
> > > > > > endpoints
> > > > > > (both implicit and explicit) as well as the isochronous (data)
> > > > > > endpoints.
> > > > > > The size of audio samples sent per USB frame (microframe) will be
> > > > > > adjusted
> > > > > > based on information received on the feedback endpoint.
> > > > >
> > > > > I think you meant "support all potential endpoint types"
> > > > >
> > > > > It's likely that some USB devices have more endpoints than what the DSP
> > > > > can handle, no?
> > > >
> > > > True, as we discussed before, we only handle the endpoints for the audio
> > > > interface.  Other endpoints, such as HID, or control is still handled by
> > > > the main processor.
> > >
> > > The number of isoc/audio endpoints can be larger than 1 per direction,
> > > it's not uncommon for a USB device to have multiple connectors on the
> > > front side for instruments, mics, monitor speakers, you name it. Just
> > > google 'motu' or 'rme usb' and you'll see examples of USB devices that
> > > are very different from plain vanilla headsets.
> >
> > Thanks for the reference.
> >
> > I tried to do some research on the RME USB audio devices, and they
> > mentioned that they do have a "class compliant mode," which is for
> > compatibility w/ Linux hosts.  I didn't see a vendor specific USB SND
> > driver matching the USB VID/PID either, so I am assuming that it uses
> > the USB SND driver as is.(and that Linux doesn't currently support their
> > vendor specific mode)  In that case, the device should conform to the
> > UAC2.0 spec (same statement seen on UAC3.0), which states in Section
> > 4.9.1 Standard AS Interface Descriptor Table 4-26:
> >
> > "4 bNumEndpoints 1 Number Number of endpoints used by this
> > interface (excluding endpoint 0). Must be
> > either 0 (no data endpoint), 1 (data
> > endpoint) or 2 (data and explicit feedback
> > endpoint)."
> >
> > So each audio streaming interface should only have 1 data and
> > potentially 1 feedback.  However, this device does expose a large number
> > of channels (I saw up to 18 channels), which the USB backend won't be
> > able to support.  I still need to check how ASoC behaves if I pass in a
> > profile that the backend can't support.
> >
> > Maybe in the non-class compliant/vendor based class driver, they have
> > the support for multiple EPs per data interface?  I don't have one of
> > these devices on hand, so I can't confirm that.
>
> Look at Figure 3-1 in the UAC2 spec, it shows it's perfectly legal to
> have multiple Audio Streaming interfaces - but one Audio Control
> interface only.
>
> The fact that there is a restriction to 1 or 2 endpoints per Audio
> Streaming interface does not really matter if in the end there are
> multiple endpoints and concurrent isoc transfers happening to/from the
> same USB device.

So the reason I wanted to mention the max number of EPs within the audio 
streaming descriptor is because the USB SND driver currently creates 
streams based off of the number of AS desc:

static int snd_usb_create_streams(struct snd_usb_audio *chip, int ctrlif)
{
...
	for (i = 0; i < assoc->bInterfaceCount; i++) {
		int intf = assoc->bFirstInterface + i;
		if (intf != ctrlif)
			snd_usb_create_stream(chip, ctrlif, intf);
	}

"assoc" is the audio control interface desc.  In the end, when userspace 
initiates a playback session, it operates on the streams created (which 
contains at max 1 isoc and 1 feedback ep)

In short, the audio DSP doesn't need to consider handling more than 1 
isoc ep (and potentially 1 feedback).  I believe that each audio stream 
creates a separate PCM device, so userspace is still free to attempt to 
activate another audio stream.  I believe # of PCM devices created 
matches the # of streams, so when userspace does activate another 
session, it would be on an entirely different substream, and can be 
handled through the USB SND (non-offload) path.  If attempted to open 
the substream used by the offload path, then we would reject is based on 
the new change.

Thanks
Wesley Cheng