[v5] media: uvcvideo: Fix bandwidth error for Alcor camera

From: Ai Chao <aichao@kylinos.cn>

The Alcor Corp. Slave camera (1b17:6684 and 2017:0011) returns a wrong
dwMaxPayloadTransferSize value for compressed formats. Valid values are
typically up to 3072 bytes per interval (for high-speed, high-bandwidth
devices), and those faulty devices request 2752512 bytes per interval.
This is a firmware issue, but the manufacturer cannot provide a fixed
firmware.

Fix this by checking the dwMaxPayloadTransferSize field, and hardcoding
a value of 1024 if it exceeds 3072 for compressed formats transferred
over isochronous endpoints. While at it, document the other quirk that
handles a bandwidth issue for uncompressed formats.

Signed-off-by: Ai Chao <aichao@kylinos.cn>
---
I have dropped the Reviewed-by tags as the patch has changed
significantly.

Ricardo, do you know if the 3072 bytes limit is fine with super-speed
devices, or does it need to be increased ?
---
 drivers/media/usb/uvc/uvc_video.c | 31 +++++++++++++++++++++++++++++++
 1 file changed, 31 insertions(+)

Hi Laurent

On Sun, 15 Jan 2023 at 22:17, Laurent Pinchart
<laurent.pinchart@ideasonboard.com> wrote:
>
> From: Ai Chao <aichao@kylinos.cn>
>
> The Alcor Corp. Slave camera (1b17:6684 and 2017:0011) returns a wrong
> dwMaxPayloadTransferSize value for compressed formats. Valid values are
> typically up to 3072 bytes per interval (for high-speed, high-bandwidth
> devices), and those faulty devices request 2752512 bytes per interval.
> This is a firmware issue, but the manufacturer cannot provide a fixed
> firmware.
>
> Fix this by checking the dwMaxPayloadTransferSize field, and hardcoding
> a value of 1024 if it exceeds 3072 for compressed formats transferred
> over isochronous endpoints. While at it, document the other quirk that
> handles a bandwidth issue for uncompressed formats.
>
> Signed-off-by: Ai Chao <aichao@kylinos.cn>
> ---
> I have dropped the Reviewed-by tags as the patch has changed
> significantly.
>
> Ricardo, do you know if the 3072 bytes limit is fine with super-speed
> devices, or does it need to be increased ?
Tried with a couple of super-speed:

If I print: ctrl->dwMaxPayloadTransferSize

[  237.269972] drivers/media/usb/uvc/uvc_video.c:239 bw 3072
[  175.761041] drivers/media/usb/uvc/uvc_video.c:239 bw 3060

Format YUYV stall when I cap the dwMaxPayloadTransferSize to 1024, but
works fine with MJPEG and even NV12

How does it sound to cap dwMaxPayloadTransferSize to 3072 for
superspeed and 1024 for high speed?

Regards!


> ---
>  drivers/media/usb/uvc/uvc_video.c | 31 +++++++++++++++++++++++++++++++
>  1 file changed, 31 insertions(+)
>
> diff --git a/drivers/media/usb/uvc/uvc_video.c b/drivers/media/usb/uvc/uvc_video.c
> index d4b023d4de7c..c6351d3b24cf 100644
> --- a/drivers/media/usb/uvc/uvc_video.c
> +++ b/drivers/media/usb/uvc/uvc_video.c
> @@ -200,6 +200,20 @@ static void uvc_fixup_video_ctrl(struct uvc_streaming *stream,
>         if ((ctrl->dwMaxPayloadTransferSize & 0xffff0000) == 0xffff0000)
>                 ctrl->dwMaxPayloadTransferSize &= ~0xffff0000;
>
> +       /*
> +        * Many devices report an incorrect dwMaxPayloadTransferSize value. The
> +        * most common issue is devices requesting the maximum possible USB
> +        * bandwidth (3072 bytes per interval for high-speed, high-bandwidth
> +        * isochronous endpoints) while they actually require less, preventing
> +        * multiple cameras from being used at the same time due to bandwidth
> +        * overallocation.
> +        *
> +        * For those devices, replace the dwMaxPayloadTransferSize value based
> +        * on an estimation calculated from the frame format and size. This is
> +        * only possible for uncompressed formats, as not enough information is
> +        * available to reliably estimate the bandwidth requirements for
> +        * compressed formats.
> +        */
>         if (!(format->flags & UVC_FMT_FLAG_COMPRESSED) &&
>             stream->dev->quirks & UVC_QUIRK_FIX_BANDWIDTH &&
>             stream->intf->num_altsetting > 1) {
> @@ -236,6 +250,23 @@ static void uvc_fixup_video_ctrl(struct uvc_streaming *stream,
>
>                 ctrl->dwMaxPayloadTransferSize = bandwidth;
>         }
> +
> +       /*
> +        * Another issue is with devices that report a transfer size that
> +        * greatly exceeds the maximum supported by any existing USB version
> +        * for isochronous transfers.  For instance, the "Slave camera" devices
> +        * from Alcor Corp. (2017:0011 and 1b17:66B8) request 2752512 bytes per
> +        * interval.
> +        *
> +        * For uncompressed formats, this can be addressed by the FIX_BANDWIDTH
> +        * quirk, but for compressed format we can't meaningfully estimate the
> +        * required bandwidth. Just hardcode it to 1024 bytes per interval,
> +        * which should be large enough for compressed formats.
> +        */
> +       if ((format->flags & UVC_FMT_FLAG_COMPRESSED) &&
> +           ctrl->dwMaxPayloadTransferSize > 3072 &&
> +           stream->intf->num_altsetting > 1)
> +               ctrl->dwMaxPayloadTransferSize = 1024;
>  }
>
>  static size_t uvc_video_ctrl_size(struct uvc_streaming *stream)
> --
> Regards,
>
> Laurent Pinchart
>

Hi Ricardo,

On Tue, Jan 17, 2023 at 03:08:07PM +0100, Ricardo Ribalda wrote:
> On Sun, 15 Jan 2023 at 22:17, Laurent Pinchart wrote:
> >
> > From: Ai Chao <aichao@kylinos.cn>
> >
> > The Alcor Corp. Slave camera (1b17:6684 and 2017:0011) returns a wrong
> > dwMaxPayloadTransferSize value for compressed formats. Valid values are
> > typically up to 3072 bytes per interval (for high-speed, high-bandwidth
> > devices), and those faulty devices request 2752512 bytes per interval.
> > This is a firmware issue, but the manufacturer cannot provide a fixed
> > firmware.
> >
> > Fix this by checking the dwMaxPayloadTransferSize field, and hardcoding
> > a value of 1024 if it exceeds 3072 for compressed formats transferred
> > over isochronous endpoints. While at it, document the other quirk that
> > handles a bandwidth issue for uncompressed formats.
> >
> > Signed-off-by: Ai Chao <aichao@kylinos.cn>
> > ---
> > I have dropped the Reviewed-by tags as the patch has changed
> > significantly.
> >
> > Ricardo, do you know if the 3072 bytes limit is fine with super-speed
> > devices, or does it need to be increased ?
> Tried with a couple of super-speed:
> 
> If I print: ctrl->dwMaxPayloadTransferSize
> 
> [  237.269972] drivers/media/usb/uvc/uvc_video.c:239 bw 3072
> [  175.761041] drivers/media/usb/uvc/uvc_video.c:239 bw 3060
> 
> Format YUYV stall when I cap the dwMaxPayloadTransferSize to 1024, but
> works fine with MJPEG and even NV12
> 
> How does it sound to cap dwMaxPayloadTransferSize to 3072 for
> superspeed and 1024 for high speed?

Won't you still run out of bandwidth when using multiple cameras
concurrently ? Is it the interval that is shorter with SS, or the
maximum bytes per interval that is larger ?

> > ---
> >  drivers/media/usb/uvc/uvc_video.c | 31 +++++++++++++++++++++++++++++++
> >  1 file changed, 31 insertions(+)
> >
> > diff --git a/drivers/media/usb/uvc/uvc_video.c b/drivers/media/usb/uvc/uvc_video.c
> > index d4b023d4de7c..c6351d3b24cf 100644
> > --- a/drivers/media/usb/uvc/uvc_video.c
> > +++ b/drivers/media/usb/uvc/uvc_video.c
> > @@ -200,6 +200,20 @@ static void uvc_fixup_video_ctrl(struct uvc_streaming *stream,
> >         if ((ctrl->dwMaxPayloadTransferSize & 0xffff0000) == 0xffff0000)
> >                 ctrl->dwMaxPayloadTransferSize &= ~0xffff0000;
> >
> > +       /*
> > +        * Many devices report an incorrect dwMaxPayloadTransferSize value. The
> > +        * most common issue is devices requesting the maximum possible USB
> > +        * bandwidth (3072 bytes per interval for high-speed, high-bandwidth
> > +        * isochronous endpoints) while they actually require less, preventing
> > +        * multiple cameras from being used at the same time due to bandwidth
> > +        * overallocation.
> > +        *
> > +        * For those devices, replace the dwMaxPayloadTransferSize value based
> > +        * on an estimation calculated from the frame format and size. This is
> > +        * only possible for uncompressed formats, as not enough information is
> > +        * available to reliably estimate the bandwidth requirements for
> > +        * compressed formats.
> > +        */
> >         if (!(format->flags & UVC_FMT_FLAG_COMPRESSED) &&
> >             stream->dev->quirks & UVC_QUIRK_FIX_BANDWIDTH &&
> >             stream->intf->num_altsetting > 1) {
> > @@ -236,6 +250,23 @@ static void uvc_fixup_video_ctrl(struct uvc_streaming *stream,
> >
> >                 ctrl->dwMaxPayloadTransferSize = bandwidth;
> >         }
> > +
> > +       /*
> > +        * Another issue is with devices that report a transfer size that
> > +        * greatly exceeds the maximum supported by any existing USB version
> > +        * for isochronous transfers.  For instance, the "Slave camera" devices
> > +        * from Alcor Corp. (2017:0011 and 1b17:66B8) request 2752512 bytes per
> > +        * interval.
> > +        *
> > +        * For uncompressed formats, this can be addressed by the FIX_BANDWIDTH
> > +        * quirk, but for compressed format we can't meaningfully estimate the
> > +        * required bandwidth. Just hardcode it to 1024 bytes per interval,
> > +        * which should be large enough for compressed formats.
> > +        */
> > +       if ((format->flags & UVC_FMT_FLAG_COMPRESSED) &&
> > +           ctrl->dwMaxPayloadTransferSize > 3072 &&
> > +           stream->intf->num_altsetting > 1)
> > +               ctrl->dwMaxPayloadTransferSize = 1024;
> >  }
> >
> >  static size_t uvc_video_ctrl_size(struct uvc_streaming *stream)

On Tue, 17 Jan 2023 at 15:18, Laurent Pinchart
<laurent.pinchart@ideasonboard.com> wrote:
>
> Hi Ricardo,
>
> On Tue, Jan 17, 2023 at 03:08:07PM +0100, Ricardo Ribalda wrote:
> > On Sun, 15 Jan 2023 at 22:17, Laurent Pinchart wrote:
> > >
> > > From: Ai Chao <aichao@kylinos.cn>
> > >
> > > The Alcor Corp. Slave camera (1b17:6684 and 2017:0011) returns a wrong
> > > dwMaxPayloadTransferSize value for compressed formats. Valid values are
> > > typically up to 3072 bytes per interval (for high-speed, high-bandwidth
> > > devices), and those faulty devices request 2752512 bytes per interval.
> > > This is a firmware issue, but the manufacturer cannot provide a fixed
> > > firmware.
> > >
> > > Fix this by checking the dwMaxPayloadTransferSize field, and hardcoding
> > > a value of 1024 if it exceeds 3072 for compressed formats transferred
> > > over isochronous endpoints. While at it, document the other quirk that
> > > handles a bandwidth issue for uncompressed formats.
> > >
> > > Signed-off-by: Ai Chao <aichao@kylinos.cn>
> > > ---
> > > I have dropped the Reviewed-by tags as the patch has changed
> > > significantly.
> > >
> > > Ricardo, do you know if the 3072 bytes limit is fine with super-speed
> > > devices, or does it need to be increased ?
> > Tried with a couple of super-speed:
> >
> > If I print: ctrl->dwMaxPayloadTransferSize
> >
> > [  237.269972] drivers/media/usb/uvc/uvc_video.c:239 bw 3072
> > [  175.761041] drivers/media/usb/uvc/uvc_video.c:239 bw 3060
> >
> > Format YUYV stall when I cap the dwMaxPayloadTransferSize to 1024, but
> > works fine with MJPEG and even NV12
> >
> > How does it sound to cap dwMaxPayloadTransferSize to 3072 for
> > superspeed and 1024 for high speed?
>
> Won't you still run out of bandwidth when using multiple cameras
> concurrently ? Is it the interval that is shorter with SS, or the
> maximum bytes per interval that is larger ?