Currently, most gadget drivers handle isoc transfers on a best effort
bases: If the request queue runs empty, then there will simply be gaps in
the isoc data stream.
The UVC gadget depends on this behaviour. It simply provides new requests
when video frames are available and assumes that they are sent as soon as
possible.
The dwc3 gadget currently works differently: It assumes that there is a
contiguous stream of requests without any gaps. If a request is too late,
then it is dropped by the hardware.
For the UVC gadget this means that a live stream stops after the first
frame because all following requests are late.
This patch changes the behaviour of the dwc3 gadget driver to match the
expectations. If a request arrives too late, then the transfer will be
restart to create the needed gap.
A request is late if:
1. There are currently no requests queued in the hardware
2. The current frame number provided by DSTS does not match the frame
number returned by the last transfer.
If this happens transfers are stopped. The following XferNotReady provides
the new correct frame number and restarts the transfer.
Signed-off-by: Michael Olbrich <[email protected]>
---
drivers/usb/dwc3/gadget.c | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c
index ac4673d91939..eb7f09929f28 100644
--- a/drivers/usb/dwc3/gadget.c
+++ b/drivers/usb/dwc3/gadget.c
@@ -1473,6 +1473,14 @@ static int __dwc3_gadget_ep_queue(struct dwc3_ep *dep, struct dwc3_request *req)
return __dwc3_gadget_start_isoc(dep);
}
}
+ if ((dep->flags & DWC3_EP_TRANSFER_STARTED) &&
+ list_empty(&dep->started_list) &&
+ ((dep->frame_number & 0x3fff) !=
+ __dwc3_gadget_get_frame(dwc))) {
+ dwc3_stop_active_transfer(dep, true, true);
+ dep->flags = DWC3_EP_ENABLED;
+ return 0;
+ }
}
return __dwc3_gadget_kick_transfer(dep, false);
--
2.20.1
Hi Michael,
Michael Olbrich wrote:
> Currently, most gadget drivers handle isoc transfers on a best effort
> bases: If the request queue runs empty, then there will simply be gaps in
> the isoc data stream.
>
> The UVC gadget depends on this behaviour. It simply provides new requests
> when video frames are available and assumes that they are sent as soon as
> possible.
>
> The dwc3 gadget currently works differently: It assumes that there is a
> contiguous stream of requests without any gaps. If a request is too late,
> then it is dropped by the hardware.
> For the UVC gadget this means that a live stream stops after the first
> frame because all following requests are late.
Can you explain little more how UVC gadget fails?
dwc3 controller expects a steady stream of data otherwise it will result
in missed_isoc status, and it should be fine as long as new requests are
queued. The controller doesn't just drop the request unless there's some
other failure.
> This patch changes the behaviour of the dwc3 gadget driver to match the
> expectations. If a request arrives too late, then the transfer will be
> restart to create the needed gap.
>
> A request is late if:
> 1. There are currently no requests queued in the hardware
> 2. The current frame number provided by DSTS does not match the frame
> number returned by the last transfer.
>
> If this happens transfers are stopped. The following XferNotReady provides
> the new correct frame number and restarts the transfer.
>
> Signed-off-by: Michael Olbrich <[email protected]>
> ---
> drivers/usb/dwc3/gadget.c | 8 ++++++++
> 1 file changed, 8 insertions(+)
>
> diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c
> index ac4673d91939..eb7f09929f28 100644
> --- a/drivers/usb/dwc3/gadget.c
> +++ b/drivers/usb/dwc3/gadget.c
> @@ -1473,6 +1473,14 @@ static int __dwc3_gadget_ep_queue(struct dwc3_ep *dep, struct dwc3_request *req)
> return __dwc3_gadget_start_isoc(dep);
> }
> }
> + if ((dep->flags & DWC3_EP_TRANSFER_STARTED) &&
> + list_empty(&dep->started_list) &&
> + ((dep->frame_number & 0x3fff) !=
> + __dwc3_gadget_get_frame(dwc))) {
> + dwc3_stop_active_transfer(dep, true, true);
> + dep->flags = DWC3_EP_ENABLED;
> + return 0;
> + }
> }
>
> return __dwc3_gadget_kick_transfer(dep, false);
BR,
Thinh
On Wed, Nov 13, 2019 at 03:55:01AM +0000, Thinh Nguyen wrote:
> Michael Olbrich wrote:
> > Currently, most gadget drivers handle isoc transfers on a best effort
> > bases: If the request queue runs empty, then there will simply be gaps in
> > the isoc data stream.
> >
> > The UVC gadget depends on this behaviour. It simply provides new requests
> > when video frames are available and assumes that they are sent as soon as
> > possible.
> >
> > The dwc3 gadget currently works differently: It assumes that there is a
> > contiguous stream of requests without any gaps. If a request is too late,
> > then it is dropped by the hardware.
> > For the UVC gadget this means that a live stream stops after the first
> > frame because all following requests are late.
>
> Can you explain little more how UVC gadget fails?
> dwc3 controller expects a steady stream of data otherwise it will result
> in missed_isoc status, and it should be fine as long as new requests are
> queued. The controller doesn't just drop the request unless there's some
> other failure.
UVC (with a live stream) does not fill the complete bandwidth of an
isochronous endpoint. Let's assume for the example that one video frame
fills 3 requests. Because it is a live stream, there will be a gap between
video frames. This is unavoidable, especially for compressed video. So the
UVC gadget will have requests for the frame numbers 1 2 3 5 6 7 9 10 11 13 14
15 and so on.
The dwc3 hardware tries to send those with frame numbers 1 2 3 4 5 6 7 8 9
10 11 12. So except for the fist few requests, all are late and result in a
missed_isoc. I tried to just ignore the missed_isoc but that did not work
for me. I only received the first frame at the other end.
Maybe I missing something here, i don't have access to the hardware
documentation, so I can only guess from the existing driver.
Regards,
Michael
> > This patch changes the behaviour of the dwc3 gadget driver to match the
> > expectations. If a request arrives too late, then the transfer will be
> > restart to create the needed gap.
> >
> > A request is late if:
> > 1. There are currently no requests queued in the hardware
> > 2. The current frame number provided by DSTS does not match the frame
> > number returned by the last transfer.
> >
> > If this happens transfers are stopped. The following XferNotReady provides
> > the new correct frame number and restarts the transfer.
> >
> > Signed-off-by: Michael Olbrich <[email protected]>
> > ---
> > drivers/usb/dwc3/gadget.c | 8 ++++++++
> > 1 file changed, 8 insertions(+)
> >
> > diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c
> > index ac4673d91939..eb7f09929f28 100644
> > --- a/drivers/usb/dwc3/gadget.c
> > +++ b/drivers/usb/dwc3/gadget.c
> > @@ -1473,6 +1473,14 @@ static int __dwc3_gadget_ep_queue(struct dwc3_ep *dep, struct dwc3_request *req)
> > return __dwc3_gadget_start_isoc(dep);
> > }
> > }
> > + if ((dep->flags & DWC3_EP_TRANSFER_STARTED) &&
> > + list_empty(&dep->started_list) &&
> > + ((dep->frame_number & 0x3fff) !=
> > + __dwc3_gadget_get_frame(dwc))) {
> > + dwc3_stop_active_transfer(dep, true, true);
> > + dep->flags = DWC3_EP_ENABLED;
> > + return 0;
> > + }
> > }
> >
> > return __dwc3_gadget_kick_transfer(dep, false);
>
> BR,
> Thinh
>
--
Pengutronix e.K. | |
Steuerwalder Str. 21 | http://www.pengutronix.de/ |
31137 Hildesheim, Germany | Phone: +49-5121-206917-0 |
Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
On Wed, 13 Nov 2019, Michael Olbrich wrote:
> On Wed, Nov 13, 2019 at 03:55:01AM +0000, Thinh Nguyen wrote:
> > Michael Olbrich wrote:
> > > Currently, most gadget drivers handle isoc transfers on a best effort
> > > bases: If the request queue runs empty, then there will simply be gaps in
> > > the isoc data stream.
> > >
> > > The UVC gadget depends on this behaviour. It simply provides new requests
> > > when video frames are available and assumes that they are sent as soon as
> > > possible.
> > >
> > > The dwc3 gadget currently works differently: It assumes that there is a
> > > contiguous stream of requests without any gaps. If a request is too late,
> > > then it is dropped by the hardware.
> > > For the UVC gadget this means that a live stream stops after the first
> > > frame because all following requests are late.
> >
> > Can you explain little more how UVC gadget fails?
> > dwc3 controller expects a steady stream of data otherwise it will result
> > in missed_isoc status, and it should be fine as long as new requests are
> > queued. The controller doesn't just drop the request unless there's some
> > other failure.
>
> UVC (with a live stream) does not fill the complete bandwidth of an
> isochronous endpoint. Let's assume for the example that one video frame
> fills 3 requests. Because it is a live stream, there will be a gap between
> video frames. This is unavoidable, especially for compressed video. So the
> UVC gadget will have requests for the frame numbers 1 2 3 5 6 7 9 10 11 13 14
> 15 and so on.
> The dwc3 hardware tries to send those with frame numbers 1 2 3 4 5 6 7 8 9
> 10 11 12. So except for the fist few requests, all are late and result in a
> missed_isoc. I tried to just ignore the missed_isoc but that did not work
> for me. I only received the first frame at the other end.
> Maybe I missing something here, i don't have access to the hardware
> documentation, so I can only guess from the existing driver.
How about changing the gadget driver instead? For frames where the UVC
gadget knows no video frame data is available (numbers 4, 8, 12, and so
on in the example above), queue a zero-length request. Then there
won't be any gaps in the isochronous packet stream.
Alan Stern
Hi Michael,
Alan Stern wrote:
> On Wed, 13 Nov 2019, Michael Olbrich wrote:
>
>> On Wed, Nov 13, 2019 at 03:55:01AM +0000, Thinh Nguyen wrote:
>>> Michael Olbrich wrote:
>>>> Currently, most gadget drivers handle isoc transfers on a best effort
>>>> bases: If the request queue runs empty, then there will simply be gaps in
>>>> the isoc data stream.
>>>>
>>>> The UVC gadget depends on this behaviour. It simply provides new requests
>>>> when video frames are available and assumes that they are sent as soon as
>>>> possible.
>>>>
>>>> The dwc3 gadget currently works differently: It assumes that there is a
>>>> contiguous stream of requests without any gaps. If a request is too late,
>>>> then it is dropped by the hardware.
>>>> For the UVC gadget this means that a live stream stops after the first
>>>> frame because all following requests are late.
>>> Can you explain little more how UVC gadget fails?
>>> dwc3 controller expects a steady stream of data otherwise it will result
>>> in missed_isoc status, and it should be fine as long as new requests are
>>> queued. The controller doesn't just drop the request unless there's some
>>> other failure.
>> UVC (with a live stream) does not fill the complete bandwidth of an
>> isochronous endpoint. Let's assume for the example that one video frame
>> fills 3 requests. Because it is a live stream, there will be a gap between
>> video frames. This is unavoidable, especially for compressed video. So the
>> UVC gadget will have requests for the frame numbers 1 2 3 5 6 7 9 10 11 13 14
>> 15 and so on.
>> The dwc3 hardware tries to send those with frame numbers 1 2 3 4 5 6 7 8 9
>> 10 11 12. So except for the fist few requests, all are late and result in a
>> missed_isoc. I tried to just ignore the missed_isoc but that did not work
>> for me. I only received the first frame at the other end.
>> Maybe I missing something here, i don't have access to the hardware
>> documentation, so I can only guess from the existing driver.
The reason I asked is because your patch doesn't seem to address the
actual issue.
For the 2 checks you do here
1. There are currently no requests queued in the hardware
2. The current frame number provided by DSTS does not match the frame
number returned by the last transfer.
For #1, it's already done in the dwc3 driver. (check
dwc3_gadget_endpoint_transfer_in_progress())
For #2, it's unlikely that DSTS current frame number will match with the
XferNotReady's frame number. So this check doesn't mean much.
So I'm still wondering how does this patch help resolve your issue.
> How about changing the gadget driver instead? For frames where the UVC
> gadget knows no video frame data is available (numbers 4, 8, 12, and so
> on in the example above), queue a zero-length request. Then there
> won't be any gaps in the isochronous packet stream.
>
> Alan Stern
What Alan suggests may work. Have you tried this?
BR,
Thinh
On Wed, Nov 13, 2019 at 07:14:59PM +0000, Thinh Nguyen wrote:
> Alan Stern wrote:
> > On Wed, 13 Nov 2019, Michael Olbrich wrote:
> >> On Wed, Nov 13, 2019 at 03:55:01AM +0000, Thinh Nguyen wrote:
> >>> Michael Olbrich wrote:
> >>>> Currently, most gadget drivers handle isoc transfers on a best effort
> >>>> bases: If the request queue runs empty, then there will simply be gaps in
> >>>> the isoc data stream.
> >>>>
> >>>> The UVC gadget depends on this behaviour. It simply provides new requests
> >>>> when video frames are available and assumes that they are sent as soon as
> >>>> possible.
> >>>>
> >>>> The dwc3 gadget currently works differently: It assumes that there is a
> >>>> contiguous stream of requests without any gaps. If a request is too late,
> >>>> then it is dropped by the hardware.
> >>>> For the UVC gadget this means that a live stream stops after the first
> >>>> frame because all following requests are late.
> >>> Can you explain little more how UVC gadget fails?
> >>> dwc3 controller expects a steady stream of data otherwise it will result
> >>> in missed_isoc status, and it should be fine as long as new requests are
> >>> queued. The controller doesn't just drop the request unless there's some
> >>> other failure.
> >> UVC (with a live stream) does not fill the complete bandwidth of an
> >> isochronous endpoint. Let's assume for the example that one video frame
> >> fills 3 requests. Because it is a live stream, there will be a gap between
> >> video frames. This is unavoidable, especially for compressed video. So the
> >> UVC gadget will have requests for the frame numbers 1 2 3 5 6 7 9 10 11 13 14
> >> 15 and so on.
> >> The dwc3 hardware tries to send those with frame numbers 1 2 3 4 5 6 7 8 9
> >> 10 11 12. So except for the fist few requests, all are late and result in a
> >> missed_isoc. I tried to just ignore the missed_isoc but that did not work
> >> for me. I only received the first frame at the other end.
> >> Maybe I missing something here, i don't have access to the hardware
> >> documentation, so I can only guess from the existing driver.
>
> The reason I asked is because your patch doesn't seem to address the
> actual issue.
>
> For the 2 checks you do here
> 1. There are currently no requests queued in the hardware
> 2. The current frame number provided by DSTS does not match the frame
> ??? number returned by the last transfer.
>
> For #1, it's already done in the dwc3 driver. (check
> dwc3_gadget_endpoint_transfer_in_progress())
But that's only after a isoc_missed occurred. What exactly does that mean?
Was the request transferred or not? My tests suggest that it was not
transferred, so I wanted to catch this before it happens.
> For #2, it's unlikely that DSTS current frame number will match with the
> XferNotReady's frame number. So this check doesn't mean much.
The frame number is also updated for each "Transfer In Progress" interrupt.
If they match, then there a new request can still be queued successfully.
Without this I got unnecessary stop/start transfers in the middle of a
video frame. But maybe something else was wrong here. I'd need to recheck.
> So I'm still wondering how does this patch help resolve your issue.
With this patch, the transfer is restarted for every video frame.
Otherwise, I just get a lot of isoc_missed and ignoring those did not help.
No valid data arrived after the first video frame.
> > How about changing the gadget driver instead? For frames where the UVC
> > gadget knows no video frame data is available (numbers 4, 8, 12, and so
> > on in the example above), queue a zero-length request. Then there
> > won't be any gaps in the isochronous packet stream.
>
> What Alan suggests may work. Have you tried this?
Yes and it works in general. There are however some problems with that
approach that I want to avoid:
1. It adds extra overhead to handle the extra zero-length request.
Especially for encoded video the available bandwidth can be quite a bit
larger that what is actually used. I want to avoid that.
2. The UVC gadget currently does no know how many zero-length request must
added. So it needs fill all available request until a new video frame
arrives. With the current 4 requests that is not a problem right now. But
that does not scale for USB3 bandwidths. So one thing that I want to do is
to queue many requests but only enable the interrupt for a few of than.
From what I can tell from the code, the gadget framework and the dwc3
driver should already support this.
This will result in extra latency. There is probably an acceptable
trade-off with an acceptable interrupt load and latency. But I would like
to avoid that if possible.
Regards,
Michael
--
Pengutronix e.K. | |
Steuerwalder Str. 21 | http://www.pengutronix.de/ |
31137 Hildesheim, Germany | Phone: +49-5121-206917-0 |
Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
Michael Olbrich wrote:
> On Wed, Nov 13, 2019 at 07:14:59PM +0000, Thinh Nguyen wrote:
>> Alan Stern wrote:
>>> On Wed, 13 Nov 2019, Michael Olbrich wrote:
>>>> On Wed, Nov 13, 2019 at 03:55:01AM +0000, Thinh Nguyen wrote:
>>>>> Michael Olbrich wrote:
>>>>>> Currently, most gadget drivers handle isoc transfers on a best effort
>>>>>> bases: If the request queue runs empty, then there will simply be gaps in
>>>>>> the isoc data stream.
>>>>>>
>>>>>> The UVC gadget depends on this behaviour. It simply provides new requests
>>>>>> when video frames are available and assumes that they are sent as soon as
>>>>>> possible.
>>>>>>
>>>>>> The dwc3 gadget currently works differently: It assumes that there is a
>>>>>> contiguous stream of requests without any gaps. If a request is too late,
>>>>>> then it is dropped by the hardware.
>>>>>> For the UVC gadget this means that a live stream stops after the first
>>>>>> frame because all following requests are late.
>>>>> Can you explain little more how UVC gadget fails?
>>>>> dwc3 controller expects a steady stream of data otherwise it will result
>>>>> in missed_isoc status, and it should be fine as long as new requests are
>>>>> queued. The controller doesn't just drop the request unless there's some
>>>>> other failure.
>>>> UVC (with a live stream) does not fill the complete bandwidth of an
>>>> isochronous endpoint. Let's assume for the example that one video frame
>>>> fills 3 requests. Because it is a live stream, there will be a gap between
>>>> video frames. This is unavoidable, especially for compressed video. So the
>>>> UVC gadget will have requests for the frame numbers 1 2 3 5 6 7 9 10 11 13 14
>>>> 15 and so on.
>>>> The dwc3 hardware tries to send those with frame numbers 1 2 3 4 5 6 7 8 9
>>>> 10 11 12. So except for the fist few requests, all are late and result in a
>>>> missed_isoc. I tried to just ignore the missed_isoc but that did not work
>>>> for me. I only received the first frame at the other end.
>>>> Maybe I missing something here, i don't have access to the hardware
>>>> documentation, so I can only guess from the existing driver.
>> The reason I asked is because your patch doesn't seem to address the
>> actual issue.
>>
>> For the 2 checks you do here
>> 1. There are currently no requests queued in the hardware
>> 2. The current frame number provided by DSTS does not match the frame
>> number returned by the last transfer.
>>
>> For #1, it's already done in the dwc3 driver. (check
>> dwc3_gadget_endpoint_transfer_in_progress())
> But that's only after a isoc_missed occurred. What exactly does that mean?
> Was the request transferred or not? My tests suggest that it was not
> transferred, so I wanted to catch this before it happens.
Missed_isoc status means that the controller did not move all the data
in an interval.
>
>> For #2, it's unlikely that DSTS current frame number will match with the
>> XferNotReady's frame number. So this check doesn't mean much.
> The frame number is also updated for each "Transfer In Progress" interrupt.
> If they match, then there a new request can still be queued successfully.
> Without this I got unnecessary stop/start transfers in the middle of a
> video frame. But maybe something else was wrong here. I'd need to recheck.
The reason they may not match is 1) the frame_number is only updated
after the software handles the XferInProgress interrupt. Depends on
system latency, that value may not be updated at the time that we check
the frame_number.
2) This check doesn't work if the service interval is greater than 1
uframe. That is, it doesn't have to match exactly the time to be
consider not late. Though, the second reason can easily be fixed.
>
>> So I'm still wondering how does this patch help resolve your issue.
> With this patch, the transfer is restarted for every video frame.
> Otherwise, I just get a lot of isoc_missed and ignoring those did not help.
> No valid data arrived after the first video frame.
>
>>> How about changing the gadget driver instead? For frames where the UVC
>>> gadget knows no video frame data is available (numbers 4, 8, 12, and so
>>> on in the example above), queue a zero-length request. Then there
>>> won't be any gaps in the isochronous packet stream.
>> What Alan suggests may work. Have you tried this?
> Yes and it works in general. There are however some problems with that
> approach that I want to avoid:
>
> 1. It adds extra overhead to handle the extra zero-length request.
> Especially for encoded video the available bandwidth can be quite a bit
> larger that what is actually used. I want to avoid that.
>
> 2. The UVC gadget currently does no know how many zero-length request must
> added. So it needs fill all available request until a new video frame
> arrives. With the current 4 requests that is not a problem right now. But
> that does not scale for USB3 bandwidths. So one thing that I want to do is
> to queue many requests but only enable the interrupt for a few of than.
> From what I can tell from the code, the gadget framework and the dwc3
> driver should already support this.
> This will result in extra latency. There is probably an acceptable
> trade-off with an acceptable interrupt load and latency. But I would like
> to avoid that if possible.
>
> Regards,
> Michael
I think I understand the problem you're trying to solve now.
The dwc3 driver does not know that there's a gap until after a new
request was queued, which then it will send an END_TRANSFER command and
dequeue all the requests to restart the transfer due to missed_isoc.
We do this because the dwc3 driver does not know whether the new request
is actually stale data, and we should not change this behavior.
Now, with UVC, it needs to communicate to the dwc3 driver that there
will be a gap after a certain request (and that the device is expecting
to send 0-length data). This is not a normal operation for isoc
transfer. You may need to introduce a new way for the function driver to
do that, possibly a new field in usb_request structure to indicate that.
However, this seems a little awkward. Maybe others can comment on this.
BR,
Thinh
On Thu, 14 Nov 2019, Thinh Nguyen wrote:
> Michael Olbrich wrote:
> >>> How about changing the gadget driver instead? For frames where the UVC
> >>> gadget knows no video frame data is available (numbers 4, 8, 12, and so
> >>> on in the example above), queue a zero-length request. Then there
> >>> won't be any gaps in the isochronous packet stream.
> >> What Alan suggests may work. Have you tried this?
> > Yes and it works in general. There are however some problems with that
> > approach that I want to avoid:
> >
> > 1. It adds extra overhead to handle the extra zero-length request.
> > Especially for encoded video the available bandwidth can be quite a bit
> > larger that what is actually used. I want to avoid that.
This comment doesn't seem to make sense. If the available bandwidth is
much _larger_ than what is actually used, what's the problem? You
don't run into difficulties until the available bandwidth is too
_small_.
The extra overhead of a zero-length request should be pretty small.
After all, the gadget expects to send a packet for every frame anyway,
more or less.
> > 2. The UVC gadget currently does no know how many zero-length request must
> > added. So it needs fill all available request until a new video frame
> > arrives. With the current 4 requests that is not a problem right now. But
> > that does not scale for USB3 bandwidths. So one thing that I want to do is
> > to queue many requests but only enable the interrupt for a few of than.
> > From what I can tell from the code, the gadget framework and the dwc3
> > driver should already support this.
> > This will result in extra latency. There is probably an acceptable
> > trade-off with an acceptable interrupt load and latency. But I would like
> > to avoid that if possible.
There are two different situations to consider:
In the middle of a video stream, latency isn't an issue.
The gadget should expect to send a new packet for each frame,
and it doesn't know what to put in that packet until it
receives the video data or it knows there won't be any data.
At the start of a video stream, latency can be an issue. But
in this situation the gadget doesn't have to send 0-length
requests until there actually is some data available.
Either way, it should be okay.
As far as interrupt load is concerned, I don't see how it relates to
the issue of sending 0-length requests.
> I think I understand the problem you're trying to solve now.
>
> The dwc3 driver does not know that there's a gap until after a new
> request was queued, which then it will send an END_TRANSFER command and
> dequeue all the requests to restart the transfer due to missed_isoc.
> We do this because the dwc3 driver does not know whether the new request
> is actually stale data, and we should not change this behavior.
>
> Now, with UVC, it needs to communicate to the dwc3 driver that there
> will be a gap after a certain request (and that the device is expecting
> to send 0-length data). This is not a normal operation for isoc
> transfer. You may need to introduce a new way for the function driver to
> do that, possibly a new field in usb_request structure to indicate that.
> However, this seems a little awkward. Maybe others can comment on this.
Note that on the host side, there is a difference between receiving
a 0-length packet and receiving no packet at all. As long as both the
host and the gadget expect the isochronous stream to be running, there
shouldn't be any gaps if you can avoid it.
Alan Stern
On Fri, Nov 15, 2019 at 04:06:10PM -0500, Alan Stern wrote:
> On Thu, 14 Nov 2019, Thinh Nguyen wrote:
>
> > Michael Olbrich wrote:
>
> > >>> How about changing the gadget driver instead? For frames where the UVC
> > >>> gadget knows no video frame data is available (numbers 4, 8, 12, and so
> > >>> on in the example above), queue a zero-length request. Then there
> > >>> won't be any gaps in the isochronous packet stream.
> > >> What Alan suggests may work. Have you tried this?
> > > Yes and it works in general. There are however some problems with that
> > > approach that I want to avoid:
> > >
> > > 1. It adds extra overhead to handle the extra zero-length request.
> > > Especially for encoded video the available bandwidth can be quite a bit
> > > larger that what is actually used. I want to avoid that.
>
> This comment doesn't seem to make sense. If the available bandwidth is
> much _larger_ than what is actually used, what's the problem? You
> don't run into difficulties until the available bandwidth is too
> _small_.
>
> The extra overhead of a zero-length request should be pretty small.
> After all, the gadget expects to send a packet for every frame anyway,
> more or less.
My current test-case is video frames with 450kB on average at 30fps. This
currently results in ~10 CPU load for the threaded interrupt handler.
At least in my test, filling the actual video data into the frame has very
little impact. So if I reserve 900kB to support occasionally larger video
frames, then I expect that this CPU load will almost double in all cases,
not just when the video frames are larger.
> > > 2. The UVC gadget currently does no know how many zero-length request must
> > > added. So it needs fill all available request until a new video frame
> > > arrives. With the current 4 requests that is not a problem right now. But
> > > that does not scale for USB3 bandwidths. So one thing that I want to do is
> > > to queue many requests but only enable the interrupt for a few of than.
> > > From what I can tell from the code, the gadget framework and the dwc3
> > > driver should already support this.
> > > This will result in extra latency. There is probably an acceptable
> > > trade-off with an acceptable interrupt load and latency. But I would like
> > > to avoid that if possible.
>
> There are two different situations to consider:
>
> In the middle of a video stream, latency isn't an issue.
> The gadget should expect to send a new packet for each frame,
> and it doesn't know what to put in that packet until it
> receives the video data or it knows there won't be any data.
>
> At the start of a video stream, latency can be an issue. But
> in this situation the gadget doesn't have to send 0-length
> requests until there actually is some data available.
>
> Either way, it should be okay.
>
> As far as interrupt load is concerned, I don't see how it relates to
> the issue of sending 0-length requests.
Maybe I don't understand, how 0-length requests work. My current
understanding is, that they are queued like any other request.
If I want to reduce the number of interrupts then I need to queue more
requests and only ask for an interrupt for some of them. This means that
potentially a lot of 0-length requests requests are queued when a new video
frame arrives and this means extra latency for the frame.
I think the worst-case latency is 2x the time between two interrupts.
So less interrupts mean more latency.
The stop/start transfer this patch implements, the video frame can be sent
immediately without any extra latency.
> > I think I understand the problem you're trying to solve now.
> >
> > The dwc3 driver does not know that there's a gap until after a new
> > request was queued, which then it will send an END_TRANSFER command and
> > dequeue all the requests to restart the transfer due to missed_isoc.
> > We do this because the dwc3 driver does not know whether the new request
> > is actually stale data, and we should not change this behavior.
> >
> > Now, with UVC, it needs to communicate to the dwc3 driver that there
> > will be a gap after a certain request (and that the device is expecting
> > to send 0-length data). This is not a normal operation for isoc
> > transfer. You may need to introduce a new way for the function driver to
> > do that, possibly a new field in usb_request structure to indicate that.
> > However, this seems a little awkward. Maybe others can comment on this.
I'm not sure how this is supposed to work. What exactly can the dwc3 driver
/ hardware do to handle a gap?
> Note that on the host side, there is a difference between receiving
> a 0-length packet and receiving no packet at all. As long as both the
> host and the gadget expect the isochronous stream to be running, there
> shouldn't be any gaps if you can avoid it.
Huh, so how is this handled on other hardware? From what I can tell the UVC
gadget works with other drivers and I've not found any special handling for
this. Is there no packet sent or are 0-length packet generated implicitly
somewhere?
Regards,
Michael
--
Pengutronix e.K. | |
Steuerwalder Str. 21 | http://www.pengutronix.de/ |
31137 Hildesheim, Germany | Phone: +49-5121-206917-0 |
Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
On Thu, 28 Nov 2019, Michael Olbrich wrote:
> On Fri, Nov 15, 2019 at 04:06:10PM -0500, Alan Stern wrote:
> > On Thu, 14 Nov 2019, Thinh Nguyen wrote:
> >
> > > Michael Olbrich wrote:
> >
> > > >>> How about changing the gadget driver instead? For frames where the UVC
> > > >>> gadget knows no video frame data is available (numbers 4, 8, 12, and so
> > > >>> on in the example above), queue a zero-length request. Then there
> > > >>> won't be any gaps in the isochronous packet stream.
> > > >> What Alan suggests may work. Have you tried this?
> > > > Yes and it works in general. There are however some problems with that
> > > > approach that I want to avoid:
> > > >
> > > > 1. It adds extra overhead to handle the extra zero-length request.
> > > > Especially for encoded video the available bandwidth can be quite a bit
> > > > larger that what is actually used. I want to avoid that.
> >
> > This comment doesn't seem to make sense. If the available bandwidth is
> > much _larger_ than what is actually used, what's the problem? You
> > don't run into difficulties until the available bandwidth is too
> > _small_.
> >
> > The extra overhead of a zero-length request should be pretty small.
> > After all, the gadget expects to send a packet for every frame anyway,
> > more or less.
>
> My current test-case is video frames with 450kB on average at 30fps. This
> currently results in ~10 CPU load for the threaded interrupt handler.
> At least in my test, filling the actual video data into the frame has very
> little impact. So if I reserve 900kB to support occasionally larger video
> frames, then I expect that this CPU load will almost double in all cases,
> not just when the video frames are larger.
This is the sort of thing you need to confirm by experimenting. It is
not at all clear that doubling the interrupt rate will also double the
CPU load, especially if half of the interrupts don't require the CPU to
do much work.
> > > > 2. The UVC gadget currently does no know how many zero-length request must
> > > > added. So it needs fill all available request until a new video frame
> > > > arrives. With the current 4 requests that is not a problem right now. But
> > > > that does not scale for USB3 bandwidths. So one thing that I want to do is
> > > > to queue many requests but only enable the interrupt for a few of than.
> > > > From what I can tell from the code, the gadget framework and the dwc3
> > > > driver should already support this.
> > > > This will result in extra latency. There is probably an acceptable
> > > > trade-off with an acceptable interrupt load and latency. But I would like
> > > > to avoid that if possible.
> >
> > There are two different situations to consider:
> >
> > In the middle of a video stream, latency isn't an issue.
> > The gadget should expect to send a new packet for each frame,
> > and it doesn't know what to put in that packet until it
> > receives the video data or it knows there won't be any data.
> >
> > At the start of a video stream, latency can be an issue. But
> > in this situation the gadget doesn't have to send 0-length
> > requests until there actually is some data available.
> >
> > Either way, it should be okay.
> >
> > As far as interrupt load is concerned, I don't see how it relates to
> > the issue of sending 0-length requests.
>
> Maybe I don't understand, how 0-length requests work. My current
> understanding is, that they are queued like any other request.
That's right.
> If I want to reduce the number of interrupts then I need to queue more
> requests and only ask for an interrupt for some of them. This means that
> potentially a lot of 0-length requests requests are queued when a new video
> frame arrives and this means extra latency for the frame.
Let's say you ask for an interrupt for only even-numbered requests.
Then there would be at most one 0-length request queued when a new
video frame arrives. Processing that 0-length request should require
very little CPU time (almost none) because it contains no data, so the
extra latency would be negligible.
> I think the worst-case latency is 2x the time between two interrupts.
> So less interrupts mean more latency.
What matters is the interrupt rate. If you double the rate at which
transfers are queued but ask for an interrupt on only half of them,
then the overall interrupt rate will remain the same and so will the
average latency.
> The stop/start transfer this patch implements, the video frame can be sent
> immediately without any extra latency.
The same would be true if you queued a request at the full isochronous
rate. If video data is present, put it in the request; if not then
queue a 0-length request.
> > > Now, with UVC, it needs to communicate to the dwc3 driver that there
> > > will be a gap after a certain request (and that the device is expecting
> > > to send 0-length data). This is not a normal operation for isoc
> > > transfer. You may need to introduce a new way for the function driver to
> > > do that, possibly a new field in usb_request structure to indicate that.
> > > However, this seems a little awkward. Maybe others can comment on this.
>
> I'm not sure how this is supposed to work. What exactly can the dwc3 driver
> / hardware do to handle a gap?
Are you talking about the driver on the gadget side or on the host
side? The rules for the host-side driver are spelled out in the
kerneldoc for usb_submit_urb(). As far as I know, there is no
equivalent set of rules for the gadget-side drivers.
> > Note that on the host side, there is a difference between receiving
> > a 0-length packet and receiving no packet at all. As long as both the
> > host and the gadget expect the isochronous stream to be running, there
> > shouldn't be any gaps if you can avoid it.
>
> Huh, so how is this handled on other hardware? From what I can tell the UVC
> gadget works with other drivers and I've not found any special handling for
> this. Is there no packet sent or are 0-length packet generated implicitly
> somewhere?
I don't know. You can ask the UVC maintainer for more information; my
guess is that it treats 0-length packets and missing packets the same.
After all, what else could it do? Either way, there is no data.
Alan Stern
On Thu, Nov 28, 2019 at 12:59:53PM -0500, Alan Stern wrote:
> On Thu, 28 Nov 2019, Michael Olbrich wrote:
> > On Fri, Nov 15, 2019 at 04:06:10PM -0500, Alan Stern wrote:
> > > On Thu, 14 Nov 2019, Thinh Nguyen wrote:
> > > > Michael Olbrich wrote:
> > > > >>> How about changing the gadget driver instead? For frames where the UVC
> > > > >>> gadget knows no video frame data is available (numbers 4, 8, 12, and so
> > > > >>> on in the example above), queue a zero-length request. Then there
> > > > >>> won't be any gaps in the isochronous packet stream.
> > > > >> What Alan suggests may work. Have you tried this?
> > > > > Yes and it works in general. There are however some problems with that
> > > > > approach that I want to avoid:
> > > > >
> > > > > 1. It adds extra overhead to handle the extra zero-length request.
> > > > > Especially for encoded video the available bandwidth can be quite a bit
> > > > > larger that what is actually used. I want to avoid that.
> > >
> > > This comment doesn't seem to make sense. If the available bandwidth is
> > > much _larger_ than what is actually used, what's the problem? You
> > > don't run into difficulties until the available bandwidth is too
> > > _small_.
> > >
> > > The extra overhead of a zero-length request should be pretty small.
> > > After all, the gadget expects to send a packet for every frame anyway,
> > > more or less.
> >
> > My current test-case is video frames with 450kB on average at 30fps. This
> > currently results in ~10 CPU load for the threaded interrupt handler.
> > At least in my test, filling the actual video data into the frame has very
> > little impact. So if I reserve 900kB to support occasionally larger video
> > frames, then I expect that this CPU load will almost double in all cases,
> > not just when the video frames are larger.
>
> This is the sort of thing you need to confirm by experimenting. It is
> not at all clear that doubling the interrupt rate will also double the
> CPU load, especially if half of the interrupts don't require the CPU to
> do much work.
As I noted before, actually filling in the video data is only a small part
of the measured CPU load. To put in in more precise numbers:
From my limited understanding, there are 8000 interrupts per second
regardless of the bandwidth (or maybe less for very low bandwidth
configurations?).
Just queuing requests without any content (so skipping the buffer handling
and 'encode()' in uvc_video_complete()) results in ~17% CPU load.
If I fill in the data for a video stream with ~1.8MB per frame and 30 fps
(and empty requests for the rest) then the CPU load goes up to ~19.5%.
This number remains the same for different bandwidths (and therefore
different request sizes and a different zero-length request percentage).
With my patches the CPU load changes as expected. The 2.5% to fill the data
remains and the rest goes down with less interrupts.
I am hoping that more batching will help here a bit. But either way the
overhead of queuing zero-length request is significant.
> > > > > 2. The UVC gadget currently does no know how many zero-length request must
> > > > > added. So it needs fill all available request until a new video frame
> > > > > arrives. With the current 4 requests that is not a problem right now. But
> > > > > that does not scale for USB3 bandwidths. So one thing that I want to do is
> > > > > to queue many requests but only enable the interrupt for a few of than.
> > > > > From what I can tell from the code, the gadget framework and the dwc3
> > > > > driver should already support this.
> > > > > This will result in extra latency. There is probably an acceptable
> > > > > trade-off with an acceptable interrupt load and latency. But I would like
> > > > > to avoid that if possible.
> > >
> > > There are two different situations to consider:
> > >
> > > In the middle of a video stream, latency isn't an issue.
> > > The gadget should expect to send a new packet for each frame,
> > > and it doesn't know what to put in that packet until it
> > > receives the video data or it knows there won't be any data.
> > >
> > > At the start of a video stream, latency can be an issue. But
> > > in this situation the gadget doesn't have to send 0-length
> > > requests until there actually is some data available.
> > >
> > > Either way, it should be okay.
> > >
> > > As far as interrupt load is concerned, I don't see how it relates to
> > > the issue of sending 0-length requests.
> >
> > Maybe I don't understand, how 0-length requests work. My current
> > understanding is, that they are queued like any other request.
>
> That's right.
>
> > If I want to reduce the number of interrupts then I need to queue more
> > requests and only ask for an interrupt for some of them. This means that
> > potentially a lot of 0-length requests requests are queued when a new video
> > frame arrives and this means extra latency for the frame.
>
> Let's say you ask for an interrupt for only even-numbered requests.
> Then there would be at most one 0-length request queued when a new
> video frame arrives. Processing that 0-length request should require
> very little CPU time (almost none) because it contains no data, so the
> extra latency would be negligible.
>
> > I think the worst-case latency is 2x the time between two interrupts.
> > So less interrupts mean more latency.
>
> What matters is the interrupt rate. If you double the rate at which
> transfers are queued but ask for an interrupt on only half of them,
> then the overall interrupt rate will remain the same and so will the
> average latency.
>
> > The stop/start transfer this patch implements, the video frame can be sent
> > immediately without any extra latency.
>
> The same would be true if you queued a request at the full isochronous
> rate. If video data is present, put it in the request; if not then
> queue a 0-length request.
My problem is this:
Let's assume a (for simplicity) that I have a video stream that fills
almost the full available bandwidth. And two reduce the bandwidth, two
interrupts will be requested for each video frame.
- When the first frame arrives, the whole frame is queued.
- When the first half of the frame is transmitted the first interrupt
arrives.
- The second frame has not yet arrived so half a frame worth of 0-length
requests are queued.
- When the second half of the frame is transmitted the second interrupt
arrives.
- The second frame has still not yet arrived so another half a frame worth
of 0-length requests are queued.
- Immediately afterwards the second frame arrives from userspace. At this
point, almost one frame worth of 0-length requests are queued so the
second frame will have an extra latency of almost one frame.
With my patch this does not happen and the transmission of the second
starts immediately.
> > > > Now, with UVC, it needs to communicate to the dwc3 driver that there
> > > > will be a gap after a certain request (and that the device is expecting
> > > > to send 0-length data). This is not a normal operation for isoc
> > > > transfer. You may need to introduce a new way for the function driver to
> > > > do that, possibly a new field in usb_request structure to indicate that.
> > > > However, this seems a little awkward. Maybe others can comment on this.
> >
> > I'm not sure how this is supposed to work. What exactly can the dwc3 driver
> > / hardware do to handle a gap?
>
> Are you talking about the driver on the gadget side or on the host
> side? The rules for the host-side driver are spelled out in the
> kerneldoc for usb_submit_urb(). As far as I know, there is no
> equivalent set of rules for the gadget-side drivers.
I'm unsure about the extra API Thinh was talking about. Right now in my
tests I'm queuing normal requests with 'request->length = 0' when no video
data is available. This has the problems described above.
The question is, can we do better than that? What could be done in the
driver if it knows that 0-length requests must be transmitted because no
new data is available?
> > > Note that on the host side, there is a difference between receiving
> > > a 0-length packet and receiving no packet at all. As long as both the
> > > host and the gadget expect the isochronous stream to be running, there
> > > shouldn't be any gaps if you can avoid it.
> >
> > Huh, so how is this handled on other hardware? From what I can tell the UVC
> > gadget works with other drivers and I've not found any special handling for
> > this. Is there no packet sent or are 0-length packet generated implicitly
> > somewhere?
>
> I don't know. You can ask the UVC maintainer for more information; my
> guess is that it treats 0-length packets and missing packets the same.
> After all, what else could it do? Either way, there is no data.
Ok.
Michael
--
Pengutronix e.K. | |
Steuerwalder Str. 21 | http://www.pengutronix.de/ |
31137 Hildesheim, Germany | Phone: +49-5121-206917-0 |
Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
On Mon, 2 Dec 2019, Michael Olbrich wrote:
> > > My current test-case is video frames with 450kB on average at 30fps. This
> > > currently results in ~10 CPU load for the threaded interrupt handler.
> > > At least in my test, filling the actual video data into the frame has very
> > > little impact. So if I reserve 900kB to support occasionally larger video
> > > frames, then I expect that this CPU load will almost double in all cases,
> > > not just when the video frames are larger.
> >
> > This is the sort of thing you need to confirm by experimenting. It is
> > not at all clear that doubling the interrupt rate will also double the
> > CPU load, especially if half of the interrupts don't require the CPU to
> > do much work.
>
> As I noted before, actually filling in the video data is only a small part
> of the measured CPU load. To put in in more precise numbers:
>
> From my limited understanding, there are 8000 interrupts per second
> regardless of the bandwidth (or maybe less for very low bandwidth
> configurations?).
> Just queuing requests without any content (so skipping the buffer handling
> and 'encode()' in uvc_video_complete()) results in ~17% CPU load.
Can you tell dwc3 to request only 4000 interrupts per second? Or 2000?
Or even 60?
> If I fill in the data for a video stream with ~1.8MB per frame and 30 fps
> (and empty requests for the rest) then the CPU load goes up to ~19.5%.
I assume you're talking about a SuperSpeed USB connection here, since
high speed can handle no more than 0.8 MB per frame at 30 fps.
> This number remains the same for different bandwidths (and therefore
> different request sizes and a different zero-length request percentage).
>
> With my patches the CPU load changes as expected. The 2.5% to fill the data
> remains and the rest goes down with less interrupts.
>
> I am hoping that more batching will help here a bit. But either way the
> overhead of queuing zero-length request is significant.
Hmmm. It may be that my thinking has been prejudiced by a
host-centered attitude, and things may work out different on the gadget
side.
There's one aspect I'm not clear on. Suppose you decide to skip
submitting requests for USB microframes 5 - 7 (at 8000 microframes
per second, as you mentioned) and then submit a request to be queued
for microframe 8. How does the UDC driver know which microframe to use
for that request? Does it always use the first available microframe?
> My problem is this:
> Let's assume a (for simplicity) that I have a video stream that fills
> almost the full available bandwidth. And two reduce the bandwidth, two
> interrupts will be requested for each video frame.
Let's assume the video frame rate is 30 fps. Since you say almost the
full available bandwidth is in use, the number of USB microframes per
video frame must be just under 8000 / 30, or around 266.
> - When the first frame arrives, the whole frame is queued.
Requiring 266 microframes, let's say starting at microframe 0.
> - When the first half of the frame is transmitted the first interrupt
> arrives.
That would be at microframe 133.
> - The second frame has not yet arrived so half a frame worth of 0-length
> requests are queued.
133 microframes, extending out to microframe 399.
> - When the second half of the frame is transmitted the second interrupt
> arrives.
At the end of microframe 266.
> - The second frame has still not yet arrived so another half a frame worth
> of 0-length requests are queued.
At this point, it has been about 266/8000 s = 33.25 ms since the first
video frame arrived. But at a rate of 30 fps, the interval between
frames should be 33.33 ms. So either something is wrong or else the
second frame is just about to arrive.
(Since the frame rate is slightly lower than the rate of transmission
of the USB data, it's inevitable that every so often you'll get an
interrupt just before a frame arrives.)
> - Immediately afterwards the second frame arrives from userspace. At this
> point, almost one frame worth of 0-length requests are queued so the
> second frame will have an extra latency of almost one frame.
I see. There is another way to approach this, however. When you get a
USB interrupt and no video data is available, nothing says you have to
queue an entire frame (or half-frame) worth of 0-length packets.
Instead, queue a request containing 2 ms (for example) of 0-length
packets. Presumably the next video frame will arrive during those 2 ms
(since it's _expected_ to arrive in no more than 0.08 ms), and from
then on you'll be okay until once again the USB data has caught up to
the video data.
On the other hand, I don't know how the UVC driver manages the
alignment between USB packets and video frames on the host end.
Perhaps it would prefer something slightly different -- that would be
another good question to ask the UVC maintainer (CC'ed).
> With my patch this does not happen and the transmission of the second
> starts immediately.
> The question is, can we do better than that? What could be done in the
> driver if it knows that 0-length requests must be transmitted because no
> new data is available?
It does seems reasonable to require that when a UDC driver encounters a
gap in an isochronous stream, it should associate the next request with
the first available microframe. If dcw3 did that then you wouldn't
have to worry about filling extra space with 0-length packets.
But if we make this change, it should be documented in an appropriate
place and it should apply to all UDC drivers -- not just dwc3.
Alan Stern
On Thu, Nov 14, 2019 at 08:11:56PM +0000, Thinh Nguyen wrote:
>Michael Olbrich wrote:
>> On Wed, Nov 13, 2019 at 07:14:59PM +0000, Thinh Nguyen wrote:
>>> Alan Stern wrote:
>>>> On Wed, 13 Nov 2019, Michael Olbrich wrote:
>>>>> On Wed, Nov 13, 2019 at 03:55:01AM +0000, Thinh Nguyen wrote:
>>>>>> Michael Olbrich wrote:
>>>>>>> Currently, most gadget drivers handle isoc transfers on a best effort
>>>>>>> bases: If the request queue runs empty, then there will simply be gaps in
>>>>>>> the isoc data stream.
>>>>>>>
>>>>>>> The UVC gadget depends on this behaviour. It simply provides new requests
>>>>>>> when video frames are available and assumes that they are sent as soon as
>>>>>>> possible.
>>>>>>>
>>>>>>> The dwc3 gadget currently works differently: It assumes that there is a
>>>>>>> contiguous stream of requests without any gaps. If a request is too late,
>>>>>>> then it is dropped by the hardware.
>>>>>>> For the UVC gadget this means that a live stream stops after the first
>>>>>>> frame because all following requests are late.
>>>>>> Can you explain little more how UVC gadget fails?
>>>>>> dwc3 controller expects a steady stream of data otherwise it will result
>>>>>> in missed_isoc status, and it should be fine as long as new requests are
>>>>>> queued. The controller doesn't just drop the request unless there's some
>>>>>> other failure.
>>>>> UVC (with a live stream) does not fill the complete bandwidth of an
>>>>> isochronous endpoint. Let's assume for the example that one video frame
>>>>> fills 3 requests. Because it is a live stream, there will be a gap between
>>>>> video frames. This is unavoidable, especially for compressed video. So the
>>>>> UVC gadget will have requests for the frame numbers 1 2 3 5 6 7 9 10 11 13 14
>>>>> 15 and so on.
>>>>> The dwc3 hardware tries to send those with frame numbers 1 2 3 4 5 6 7 8 9
>>>>> 10 11 12. So except for the fist few requests, all are late and result in a
>>>>> missed_isoc. I tried to just ignore the missed_isoc but that did not work
>>>>> for me. I only received the first frame at the other end.
>>>>> Maybe I missing something here, i don't have access to the hardware
>>>>> documentation, so I can only guess from the existing driver.
>>> The reason I asked is because your patch doesn't seem to address the
>>> actual issue.
>>>
>>> For the 2 checks you do here
>>> 1. There are currently no requests queued in the hardware
>>> 2. The current frame number provided by DSTS does not match the frame
>>> ??? number returned by the last transfer.
>>>
>>> For #1, it's already done in the dwc3 driver. (check
>>> dwc3_gadget_endpoint_transfer_in_progress())
>> But that's only after a isoc_missed occurred. What exactly does that mean?
>> Was the request transferred or not? My tests suggest that it was not
>> transferred, so I wanted to catch this before it happens.
>
>Missed_isoc status means that the controller did not move all the data
>in an interval.
I read in some Processor documentation that in case the host tries to
fetch data from the client and no active TRB (HWO=1) is available the
XferInProgress Interrupt will be produced, with the missed status set.
This is done because the hardware will produce zero length packets
on its own, to keep the stream running.
>>> For #2, it's unlikely that DSTS current frame number will match with the
>>> XferNotReady's frame number. So this check doesn't mean much.
>> The frame number is also updated for each "Transfer In Progress" interrupt.
>> If they match, then there a new request can still be queued successfully.
>> Without this I got unnecessary stop/start transfers in the middle of a
>> video frame. But maybe something else was wrong here. I'd need to recheck.
>
>The reason they may not match is 1) the frame_number is only updated
>after the software handles the XferInProgress interrupt. Depends on
>system latency, that value may not be updated at the time that we check
>the frame_number.
>2) This check doesn't work if the service interval is greater than 1
>uframe. That is, it doesn't have to match exactly the time to be
>consider not late. Though, the second reason can easily be fixed.
In the empty trb case, after the Hardware has send enough zero packets this
active transfer has to be stopped with endtransfer cmd. Because every next
update transfer on that active transfer will likely lead to further missed
transfers, as the newly updated trb will be handled to late anyway.
The odd thing here is, that I don't see the refered XferInProgress
Interrupts with the missed event, when the started_list is empty.
But this would be the only case to fall into this condition and handle it
properly. Like alredy assumed in the following code:
static void dwc3_gadget_endpoint_transfer_in_progress(struct dwc3_ep *dep,
const struct dwc3_event_depevt *event)
{
...
if (event->status & DEPEVT_STATUS_MISSED_ISOC) {
status = -EXDEV;
if (list_empty(&dep->started_list))
stop = true;
}
...
if (stop)
dwc3_stop_active_transfer(dep, true, true);
...
}
In fact I did sometimes see these XferInProgress Interrupts on empty trb queue
after I stoped the tansfer when the started_list was empty right after
ep_cleanup_completed_requests has moved all trbs out of the queue.
These Interrupts appeared right after the ENDTRANSFER cmd was send. (But I
could no verify this every time)
Anyways in that case these Interrupts are not useful anymore, as I already
implied the same stop, with ENDTRANSFER after we know that there are no other
trbs in the chain.
Regards,
Michael
--
Pengutronix e.K. | |
Steuerwalder Str. 21 | http://www.pengutronix.de/ |
31137 Hildesheim, Germany | Phone: +49-5121-206917-0 |
Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
Hi,
Michael Grzeschik wrote:
> On Thu, Nov 14, 2019 at 08:11:56PM +0000, Thinh Nguyen wrote:
>> Michael Olbrich wrote:
>>> On Wed, Nov 13, 2019 at 07:14:59PM +0000, Thinh Nguyen wrote:
>>>> Alan Stern wrote:
>>>>> On Wed, 13 Nov 2019, Michael Olbrich wrote:
>>>>>> On Wed, Nov 13, 2019 at 03:55:01AM +0000, Thinh Nguyen wrote:
>>>>>>> Michael Olbrich wrote:
>>>>>>>> Currently, most gadget drivers handle isoc transfers on a best
>>>>>>>> effort
>>>>>>>> bases: If the request queue runs empty, then there will simply
>>>>>>>> be gaps in
>>>>>>>> the isoc data stream.
>>>>>>>>
>>>>>>>> The UVC gadget depends on this behaviour. It simply provides
>>>>>>>> new requests
>>>>>>>> when video frames are available and assumes that they are sent
>>>>>>>> as soon as
>>>>>>>> possible.
>>>>>>>>
>>>>>>>> The dwc3 gadget currently works differently: It assumes that
>>>>>>>> there is a
>>>>>>>> contiguous stream of requests without any gaps. If a request is
>>>>>>>> too late,
>>>>>>>> then it is dropped by the hardware.
>>>>>>>> For the UVC gadget this means that a live stream stops after
>>>>>>>> the first
>>>>>>>> frame because all following requests are late.
>>>>>>> Can you explain little more how UVC gadget fails?
>>>>>>> dwc3 controller expects a steady stream of data otherwise it
>>>>>>> will result
>>>>>>> in missed_isoc status, and it should be fine as long as new
>>>>>>> requests are
>>>>>>> queued. The controller doesn't just drop the request unless
>>>>>>> there's some
>>>>>>> other failure.
>>>>>> UVC (with a live stream) does not fill the complete bandwidth of an
>>>>>> isochronous endpoint. Let's assume for the example that one video
>>>>>> frame
>>>>>> fills 3 requests. Because it is a live stream, there will be a
>>>>>> gap between
>>>>>> video frames. This is unavoidable, especially for compressed
>>>>>> video. So the
>>>>>> UVC gadget will have requests for the frame numbers 1 2 3 5 6 7 9
>>>>>> 10 11 13 14
>>>>>> 15 and so on.
>>>>>> The dwc3 hardware tries to send those with frame numbers 1 2 3 4
>>>>>> 5 6 7 8 9
>>>>>> 10 11 12. So except for the fist few requests, all are late and
>>>>>> result in a
>>>>>> missed_isoc. I tried to just ignore the missed_isoc but that did
>>>>>> not work
>>>>>> for me. I only received the first frame at the other end.
>>>>>> Maybe I missing something here, i don't have access to the hardware
>>>>>> documentation, so I can only guess from the existing driver.
>>>> The reason I asked is because your patch doesn't seem to address the
>>>> actual issue.
>>>>
>>>> For the 2 checks you do here
>>>> 1. There are currently no requests queued in the hardware
>>>> 2. The current frame number provided by DSTS does not match the frame
>>>> ? ??? number returned by the last transfer.
>>>>
>>>> For #1, it's already done in the dwc3 driver. (check
>>>> dwc3_gadget_endpoint_transfer_in_progress())
>>> But that's only after a isoc_missed occurred. What exactly does that
>>> mean?
>>> Was the request transferred or not? My tests suggest that it was not
>>> transferred, so I wanted to catch this before it happens.
>>
>> Missed_isoc status means that the controller did not move all the data
>> in an interval.
>
> I read in some Processor documentation that in case the host tries to
> fetch data from the client and no active TRB (HWO=1) is available the
> XferInProgress Interrupt will be produced, with the missed status set.
> This is done because the hardware will produce zero length packets
> on its own, to keep the stream running.
The controller only generates XferInProgress if it had processed a TRB
(with specific control bits). For IN direction, if the controller is
starved of TRB, it will send a ZLP if the host requests for data.
>
>>>> For #2, it's unlikely that DSTS current frame number will match
>>>> with the
>>>> XferNotReady's frame number. So this check doesn't mean much.
>>> The frame number is also updated for each "Transfer In Progress"
>>> interrupt.
>>> If they match, then there a new request can still be queued
>>> successfully.
>>> Without this I got unnecessary stop/start transfers in the middle of a
>>> video frame. But maybe something else was wrong here. I'd need to
>>> recheck.
>>
>> The reason they may not match is 1) the frame_number is only updated
>> after the software handles the XferInProgress interrupt. Depends on
>> system latency, that value may not be updated at the time that we check
>> the frame_number.
>> 2) This check doesn't work if the service interval is greater than 1
>> uframe. That is, it doesn't have to match exactly the time to be
>> consider not late. Though, the second reason can easily be fixed.
>
> In the empty trb case, after the Hardware has send enough zero packets
> this
> active transfer has to be stopped with endtransfer cmd. Because every
> next
> update transfer on that active transfer will likely lead to further
> missed
> transfers, as the newly updated trb will be handled to late anyway.
The controller is expecting the function driver to feed TRBs to the
controller for every interval. If it's late, then the controller will
consider that data "missed_isoc".
In your case, the UVC driver seems to queue requests to the controller
driver as if it is bulk requests, and the UVC expects those data to go
out at the time it queues. To achieve what UVC needs, then you may want
to issue END_TRANSFER command before the next burst of data. This way,
the controller can restart the isoc endpoint and not consider the next
video frame data late. There are some corner cases that you need to
watch out for. If you're going for this route, we can look further.
Also, you'd need provide a way for the UVC to communicate to the dwc3 to
let it know to expect the next burst of data.
>
>
> The odd thing here is, that I don't see the refered XferInProgress
> Interrupts with the missed event, when the started_list is empty.
See my first comment.
>
> But this would be the only case to fall into this condition and handle it
> properly. Like alredy assumed in the following code:
>
> static void dwc3_gadget_endpoint_transfer_in_progress(struct dwc3_ep
> *dep,
> ???????????? const struct dwc3_event_depevt *event)
> {
> ...
>
> ?????? if (event->status & DEPEVT_STATUS_MISSED_ISOC) {
> ?????????????? status = -EXDEV;
>
> ?????????????? if (list_empty(&dep->started_list))
> ?????????????????????? stop = true;
> ?????? }
>
> ...
>
> ?????? if (stop)
> ?????????????? dwc3_stop_active_transfer(dep, true, true);
> ...
> }
>
> In fact I did sometimes see these XferInProgress Interrupts on empty
> trb queue
> after I stoped the tansfer when the started_list was empty right after
> ep_cleanup_completed_requests has moved all trbs out of the queue.
>
> These Interrupts appeared right after the ENDTRANSFER cmd was send.
> (But I
> could no verify this every time)
If END_TRANSFER command completes, then you should not see
XferInProgress event. The next event should ber XferNotReady.
>
> Anyways in that case these Interrupts are not useful anymore, as I
> already
> implied the same stop, with ENDTRANSFER after we know that there are
> no other
> trbs in the chain.
>
Just curious, do you know if there's a reason for UVC to behave this
way? Seems like what it's trying to do is more for bulk. Maybe it wants
bandwidth priority perhaps?
BR,
Thinh
On Fri, Apr 10, 2020 at 01:09:23AM +0000, Thinh Nguyen wrote:
>Hi,
>
>Michael Grzeschik wrote:
>> On Thu, Nov 14, 2019 at 08:11:56PM +0000, Thinh Nguyen wrote:
>>> Michael Olbrich wrote:
>>>> On Wed, Nov 13, 2019 at 07:14:59PM +0000, Thinh Nguyen wrote:
>>>>> Alan Stern wrote:
>>>>>> On Wed, 13 Nov 2019, Michael Olbrich wrote:
>>>>>>> On Wed, Nov 13, 2019 at 03:55:01AM +0000, Thinh Nguyen wrote:
>>>>>>>> Michael Olbrich wrote:
>>>>>>>>> Currently, most gadget drivers handle isoc transfers on a best
>>>>>>>>> effort
>>>>>>>>> bases: If the request queue runs empty, then there will simply
>>>>>>>>> be gaps in
>>>>>>>>> the isoc data stream.
>>>>>>>>>
>>>>>>>>> The UVC gadget depends on this behaviour. It simply provides
>>>>>>>>> new requests
>>>>>>>>> when video frames are available and assumes that they are sent
>>>>>>>>> as soon as
>>>>>>>>> possible.
>>>>>>>>>
>>>>>>>>> The dwc3 gadget currently works differently: It assumes that
>>>>>>>>> there is a
>>>>>>>>> contiguous stream of requests without any gaps. If a request is
>>>>>>>>> too late,
>>>>>>>>> then it is dropped by the hardware.
>>>>>>>>> For the UVC gadget this means that a live stream stops after
>>>>>>>>> the first
>>>>>>>>> frame because all following requests are late.
>>>>>>>> Can you explain little more how UVC gadget fails?
>>>>>>>> dwc3 controller expects a steady stream of data otherwise it
>>>>>>>> will result
>>>>>>>> in missed_isoc status, and it should be fine as long as new
>>>>>>>> requests are
>>>>>>>> queued. The controller doesn't just drop the request unless
>>>>>>>> there's some
>>>>>>>> other failure.
>>>>>>> UVC (with a live stream) does not fill the complete bandwidth of an
>>>>>>> isochronous endpoint. Let's assume for the example that one video
>>>>>>> frame
>>>>>>> fills 3 requests. Because it is a live stream, there will be a
>>>>>>> gap between
>>>>>>> video frames. This is unavoidable, especially for compressed
>>>>>>> video. So the
>>>>>>> UVC gadget will have requests for the frame numbers 1 2 3 5 6 7 9
>>>>>>> 10 11 13 14
>>>>>>> 15 and so on.
>>>>>>> The dwc3 hardware tries to send those with frame numbers 1 2 3 4
>>>>>>> 5 6 7 8 9
>>>>>>> 10 11 12. So except for the fist few requests, all are late and
>>>>>>> result in a
>>>>>>> missed_isoc. I tried to just ignore the missed_isoc but that did
>>>>>>> not work
>>>>>>> for me. I only received the first frame at the other end.
>>>>>>> Maybe I missing something here, i don't have access to the hardware
>>>>>>> documentation, so I can only guess from the existing driver.
>>>>> The reason I asked is because your patch doesn't seem to address the
>>>>> actual issue.
>>>>>
>>>>> For the 2 checks you do here
>>>>> 1. There are currently no requests queued in the hardware
>>>>> 2. The current frame number provided by DSTS does not match the frame
>>>>> ? ??? number returned by the last transfer.
>>>>>
>>>>> For #1, it's already done in the dwc3 driver. (check
>>>>> dwc3_gadget_endpoint_transfer_in_progress())
>>>> But that's only after a isoc_missed occurred. What exactly does that
>>>> mean?
>>>> Was the request transferred or not? My tests suggest that it was not
>>>> transferred, so I wanted to catch this before it happens.
>>>
>>> Missed_isoc status means that the controller did not move all the data
>>> in an interval.
>>
>> I read in some Processor documentation that in case the host tries to
>> fetch data from the client and no active TRB (HWO=1) is available the
>> XferInProgress Interrupt will be produced, with the missed status set.
>> This is done because the hardware will produce zero length packets
>> on its own, to keep the stream running.
>
>The controller only generates XferInProgress if it had processed a TRB
>(with specific control bits). For IN direction, if the controller is
>starved of TRB, it will send a ZLP if the host requests for data.
Which control bits are those? ISOC-First, Chain and Last bits?
I see the Scatter-Gather preparation is using these pattern.
>>>>> For #2, it's unlikely that DSTS current frame number will match
>>>>> with the
>>>>> XferNotReady's frame number. So this check doesn't mean much.
>>>> The frame number is also updated for each "Transfer In Progress"
>>>> interrupt.
>>>> If they match, then there a new request can still be queued
>>>> successfully.
>>>> Without this I got unnecessary stop/start transfers in the middle of a
>>>> video frame. But maybe something else was wrong here. I'd need to
>>>> recheck.
>>>
>>> The reason they may not match is 1) the frame_number is only updated
>>> after the software handles the XferInProgress interrupt. Depends on
>>> system latency, that value may not be updated at the time that we check
>>> the frame_number.
>>> 2) This check doesn't work if the service interval is greater than 1
>>> uframe. That is, it doesn't have to match exactly the time to be
>>> consider not late. Though, the second reason can easily be fixed.
>>
>> In the empty trb case, after the Hardware has send enough zero packets
>> this
>> active transfer has to be stopped with endtransfer cmd. Because every
>> next
>> update transfer on that active transfer will likely lead to further
>> missed
>> transfers, as the newly updated trb will be handled to late anyway.
>
>The controller is expecting the function driver to feed TRBs to the
>controller for every interval. If it's late, then the controller will
>consider that data "missed_isoc".
>
>In your case, the UVC driver seems to queue requests to the controller
>driver as if it is bulk requests, and the UVC expects those data to go
>out at the time it queues. To achieve what UVC needs, then you may want
>to issue END_TRANSFER command before the next burst of data. This way,
>the controller can restart the isoc endpoint and not consider the next
>video frame data late. There are some corner cases that you need to
>watch out for. If you're going for this route, we can look further.
Right, for now the drivers is doing:
- Strart Transfer (with the right frame number) once.
- Update Transfer On every ep_queue with the corresponding TRB
setting CHN = 0, IOC = 1, First-ISOC = 1
- End Transfer is somehow not handled right for this case.
See my first comment. I think dwc3_prepare_one_trb_sg does the proper chain
handling already.
>Also, you'd need provide a way for the UVC to communicate to the dwc3 to
>let it know to expect the next burst of data.
Can the UVC not just enqueue one big sg-request, which will represent one burst
and not one TRB. Also that is what the SG code already seem to handle right.
>> The odd thing here is, that I don't see the refered XferInProgress
>> Interrupts with the missed event, when the started_list is empty.
>
>See my first comment.
>
>>
>> But this would be the only case to fall into this condition and handle it
>> properly. Like alredy assumed in the following code:
>>
>> static void dwc3_gadget_endpoint_transfer_in_progress(struct dwc3_ep
>> *dep,
>> ???????????? const struct dwc3_event_depevt *event)
>> {
>> ...
>>
>> ?????? if (event->status & DEPEVT_STATUS_MISSED_ISOC) {
>> ?????????????? status = -EXDEV;
>>
>> ?????????????? if (list_empty(&dep->started_list))
>> ?????????????????????? stop = true;
>> ?????? }
>>
>> ...
>>
>> ?????? if (stop)
>> ?????????????? dwc3_stop_active_transfer(dep, true, true);
>> ...
>> }
>>
>> In fact I did sometimes see these XferInProgress Interrupts on empty
>> trb queue
>> after I stoped the tansfer when the started_list was empty right after
>> ep_cleanup_completed_requests has moved all trbs out of the queue.
>>
>> These Interrupts appeared right after the ENDTRANSFER cmd was send.
>> (But I
>> could no verify this every time)
>
>If END_TRANSFER command completes, then you should not see
>XferInProgress event. The next event should ber XferNotReady.
Right. This also stops, after the Command Complete Interrupt arrives.
>> Anyways in that case these Interrupts are not useful anymore, as I
>> already
>> implied the same stop, with ENDTRANSFER after we know that there are
>> no other
>> trbs in the chain.
>>
>
>Just curious, do you know if there's a reason for UVC to behave this
>way? Seems like what it's trying to do is more for bulk. Maybe it wants
>bandwidth priority perhaps?
I don't know, probably it was more likely for USB 2.0 controllers to be handled
this way.
As mentioned the current uvc code is also working when we add this changes.
diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c
index ec357f64f319..a5dc44f2e9d8 100644
--- a/drivers/usb/dwc3/gadget.c
+++ b/drivers/usb/dwc3/gadget.c
@@ -2629,6 +2629,9 @@ static void dwc3_gadget_endpoint_transfer_in_progress(struct dwc3_ep *dep,
dwc3_gadget_ep_cleanup_completed_requests(dep, event, status);
+ if (list_empty(&dep->started_list))
+ stop = true;
+
if (stop)
dwc3_stop_active_transfer(dep, true, true);
else if (dwc3_gadget_ep_should_continue(dep))
diff --git a/drivers/usb/gadget/function/uvc_video.c b/drivers/usb/gadget/function/uvc_video.c
index da6ba8ba4bca..a3dac5d91aae 100644
--- a/drivers/usb/gadget/function/uvc_video.c
+++ b/drivers/usb/gadget/function/uvc_video.c
@@ -183,6 +183,7 @@ uvc_video_complete(struct usb_ep *ep, struct usb_request *req)
switch (req->status) {
case 0:
+ case -EXDEV: /* we ignore missed transfers */
break;
case -ESHUTDOWN: /* disconnect from host. */
Regards,
Michael
--
Pengutronix e.K. | |
Steuerwalder Str. 21 | http://www.pengutronix.de/ |
31137 Hildesheim, Germany | Phone: +49-5121-206917-0 |
Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
Hi,
Michael Grzeschik wrote:
> On Fri, Apr 10, 2020 at 01:09:23AM +0000, Thinh Nguyen wrote:
>> Hi,
>>
>> Michael Grzeschik wrote:
>>> On Thu, Nov 14, 2019 at 08:11:56PM +0000, Thinh Nguyen wrote:
>>>> Michael Olbrich wrote:
>>>>> On Wed, Nov 13, 2019 at 07:14:59PM +0000, Thinh Nguyen wrote:
>>>>>> Alan Stern wrote:
>>>>>>> On Wed, 13 Nov 2019, Michael Olbrich wrote:
>>>>>>>> On Wed, Nov 13, 2019 at 03:55:01AM +0000, Thinh Nguyen wrote:
>>>>>>>>> Michael Olbrich wrote:
>>>>>>>>>> Currently, most gadget drivers handle isoc transfers on a best
>>>>>>>>>> effort
>>>>>>>>>> bases: If the request queue runs empty, then there will simply
>>>>>>>>>> be gaps in
>>>>>>>>>> the isoc data stream.
>>>>>>>>>>
>>>>>>>>>> The UVC gadget depends on this behaviour. It simply provides
>>>>>>>>>> new requests
>>>>>>>>>> when video frames are available and assumes that they are sent
>>>>>>>>>> as soon as
>>>>>>>>>> possible.
>>>>>>>>>>
>>>>>>>>>> The dwc3 gadget currently works differently: It assumes that
>>>>>>>>>> there is a
>>>>>>>>>> contiguous stream of requests without any gaps. If a request is
>>>>>>>>>> too late,
>>>>>>>>>> then it is dropped by the hardware.
>>>>>>>>>> For the UVC gadget this means that a live stream stops after
>>>>>>>>>> the first
>>>>>>>>>> frame because all following requests are late.
>>>>>>>>> Can you explain little more how UVC gadget fails?
>>>>>>>>> dwc3 controller expects a steady stream of data otherwise it
>>>>>>>>> will result
>>>>>>>>> in missed_isoc status, and it should be fine as long as new
>>>>>>>>> requests are
>>>>>>>>> queued. The controller doesn't just drop the request unless
>>>>>>>>> there's some
>>>>>>>>> other failure.
>>>>>>>> UVC (with a live stream) does not fill the complete bandwidth
>>>>>>>> of an
>>>>>>>> isochronous endpoint. Let's assume for the example that one video
>>>>>>>> frame
>>>>>>>> fills 3 requests. Because it is a live stream, there will be a
>>>>>>>> gap between
>>>>>>>> video frames. This is unavoidable, especially for compressed
>>>>>>>> video. So the
>>>>>>>> UVC gadget will have requests for the frame numbers 1 2 3 5 6 7 9
>>>>>>>> 10 11 13 14
>>>>>>>> 15 and so on.
>>>>>>>> The dwc3 hardware tries to send those with frame numbers 1 2 3 4
>>>>>>>> 5 6 7 8 9
>>>>>>>> 10 11 12. So except for the fist few requests, all are late and
>>>>>>>> result in a
>>>>>>>> missed_isoc. I tried to just ignore the missed_isoc but that did
>>>>>>>> not work
>>>>>>>> for me. I only received the first frame at the other end.
>>>>>>>> Maybe I missing something here, i don't have access to the
>>>>>>>> hardware
>>>>>>>> documentation, so I can only guess from the existing driver.
>>>>>> The reason I asked is because your patch doesn't seem to address the
>>>>>> actual issue.
>>>>>>
>>>>>> For the 2 checks you do here
>>>>>> 1. There are currently no requests queued in the hardware
>>>>>> 2. The current frame number provided by DSTS does not match the
>>>>>> frame
>>>>>> ? ??? number returned by the last transfer.
>>>>>>
>>>>>> For #1, it's already done in the dwc3 driver. (check
>>>>>> dwc3_gadget_endpoint_transfer_in_progress())
>>>>> But that's only after a isoc_missed occurred. What exactly does that
>>>>> mean?
>>>>> Was the request transferred or not? My tests suggest that it was not
>>>>> transferred, so I wanted to catch this before it happens.
>>>>
>>>> Missed_isoc status means that the controller did not move all the data
>>>> in an interval.
>>>
>>> I read in some Processor documentation that in case the host tries to
>>> fetch data from the client and no active TRB (HWO=1) is available the
>>> XferInProgress Interrupt will be produced, with the missed status set.
>>> This is done because the hardware will produce zero length packets
>>> on its own, to keep the stream running.
>>
>> The controller only generates XferInProgress if it had processed a TRB
>> (with specific control bits). For IN direction, if the controller is
>> starved of TRB, it will send a ZLP if the host requests for data.
>
> Which control bits are those? ISOC-First, Chain and Last bits?
>
> I see the Scatter-Gather preparation is using these pattern.
The IOC bit. You can check the programming guide for more detail on how
to setup the TRBs, but what we have in dwc3 is fine.
>
>>>>>> For #2, it's unlikely that DSTS current frame number will match
>>>>>> with the
>>>>>> XferNotReady's frame number. So this check doesn't mean much.
>>>>> The frame number is also updated for each "Transfer In Progress"
>>>>> interrupt.
>>>>> If they match, then there a new request can still be queued
>>>>> successfully.
>>>>> Without this I got unnecessary stop/start transfers in the middle
>>>>> of a
>>>>> video frame. But maybe something else was wrong here. I'd need to
>>>>> recheck.
>>>>
>>>> The reason they may not match is 1) the frame_number is only updated
>>>> after the software handles the XferInProgress interrupt. Depends on
>>>> system latency, that value may not be updated at the time that we
>>>> check
>>>> the frame_number.
>>>> 2) This check doesn't work if the service interval is greater than 1
>>>> uframe. That is, it doesn't have to match exactly the time to be
>>>> consider not late. Though, the second reason can easily be fixed.
>>>
>>> In the empty trb case, after the Hardware has send enough zero packets
>>> this
>>> active transfer has to be stopped with endtransfer cmd. Because every
>>> next
>>> update transfer on that active transfer will likely lead to further
>>> missed
>>> transfers, as the newly updated trb will be handled to late anyway.
>>
>> The controller is expecting the function driver to feed TRBs to the
>> controller for every interval. If it's late, then the controller will
>> consider that data "missed_isoc".
>>
>> In your case, the UVC driver seems to queue requests to the controller
>> driver as if it is bulk requests, and the UVC expects those data to go
>> out at the time it queues. To achieve what UVC needs, then you may want
>> to issue END_TRANSFER command before the next burst of data. This way,
>> the controller can restart the isoc endpoint and not consider the next
>> video frame data late. There are some corner cases that you need to
>> watch out for. If you're going for this route, we can look further.
>
> Right, for now the drivers is doing:
>
> - Strart Transfer (with the right frame number) once.
>
> - Update Transfer On every ep_queue with the corresponding TRB
> ?setting CHN = 0, IOC = 1, First-ISOC = 1
>
> - End Transfer is somehow not handled right for this case.
>
> See my first comment. I think dwc3_prepare_one_trb_sg does the proper
> chain
> handling already.
>
>> Also, you'd need provide a way for the UVC to communicate to the dwc3 to
>> let it know to expect the next burst of data.
>
> Can the UVC not just enqueue one big sg-request, which will represent
> one burst
> and not one TRB. Also that is? what the SG code already seem to handle
> right.
Do you need SG? What size does video class driver setup its isoc URB? If
it's superspeed, max is 48K, and depending on the type of platform
you're running UVC on, you can setup a single 48K buffer per request if
you want to do that. However, it's probably not a good idea since many
host controllers can't even handle even close to 48K/uframe.
What I was saying is if UVC knows when to wait for the next video data,
it can tell dwc3 to stop the isoc endpoint before queuing the next video
data in a set of requests. If UVC doesn't know that, then it needs to
tell dwc3 to change its handling of isoc requests.
>
>>> The odd thing here is, that I don't see the refered XferInProgress
>>> Interrupts with the missed event, when the started_list is empty.
>>
>> See my first comment.
>>
>>>
>>> But this would be the only case to fall into this condition and
>>> handle it
>>> properly. Like alredy assumed in the following code:
>>>
>>> static void dwc3_gadget_endpoint_transfer_in_progress(struct dwc3_ep
>>> *dep,
>>> ???????????? const struct dwc3_event_depevt *event)
>>> {
>>> ...
>>>
>>> ?????? if (event->status & DEPEVT_STATUS_MISSED_ISOC) {
>>> ?????????????? status = -EXDEV;
>>>
>>> ?????????????? if (list_empty(&dep->started_list))
>>> ?????????????????????? stop = true;
>>> ?????? }
>>>
>>> ...
>>>
>>> ?????? if (stop)
>>> ?????????????? dwc3_stop_active_transfer(dep, true, true);
>>> ...
>>> }
>>>
>>> In fact I did sometimes see these XferInProgress Interrupts on empty
>>> trb queue
>>> after I stoped the tansfer when the started_list was empty right after
>>> ep_cleanup_completed_requests has moved all trbs out of the queue.
>>>
>>> These Interrupts appeared right after the ENDTRANSFER cmd was send.
>>> (But I
>>> could no verify this every time)
>>
>> If END_TRANSFER command completes, then you should not see
>> XferInProgress event. The next event should ber XferNotReady.
>
> Right. This also stops, after the Command Complete Interrupt arrives.
>
>>> Anyways in that case these Interrupts are not useful anymore, as I
>>> already
>>> implied the same stop, with ENDTRANSFER after we know that there are
>>> no other
>>> trbs in the chain.
>>>
>>
>> Just curious, do you know if there's a reason for UVC to behave this
>> way? Seems like what it's trying to do is more for bulk. Maybe it wants
>> bandwidth priority perhaps?
>
> I don't know, probably it was more likely for USB 2.0 controllers to
> be handled
> this way.
>
> As mentioned the current uvc code is also working when we add this
> changes.
>
> diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c
> index ec357f64f319..a5dc44f2e9d8 100644
> --- a/drivers/usb/dwc3/gadget.c
> +++ b/drivers/usb/dwc3/gadget.c
> @@ -2629,6 +2629,9 @@ static void
> dwc3_gadget_endpoint_transfer_in_progress(struct dwc3_ep *dep,
>
> ?????? dwc3_gadget_ep_cleanup_completed_requests(dep, event, status);
>
> +?????? if (list_empty(&dep->started_list))
> +?????????????? stop = true;
> +
You should check the pending list too and either drop them or prepare
those requests (maybe too late). This happens when there's no available
TRB but UVC queues more requests.
Also, make sure this change only applies for isoc.
This may work. Just keep in mind that the timing parameter of the
XferNotReady will be expired by the time the UVC queues the next isoc
request. (Maybe that's why you tried to check for the current frame
number via DSTS instead in your first patch?).
With the new changes in Felipe testing/next branch, this may be ok.
> ?????? if (stop)
> ?????????????? dwc3_stop_active_transfer(dep, true, true);
> ?????? else if (dwc3_gadget_ep_should_continue(dep))
>
> diff --git a/drivers/usb/gadget/function/uvc_video.c
> b/drivers/usb/gadget/function/uvc_video.c
> index da6ba8ba4bca..a3dac5d91aae 100644
> --- a/drivers/usb/gadget/function/uvc_video.c
> +++ b/drivers/usb/gadget/function/uvc_video.c
> @@ -183,6 +183,7 @@ uvc_video_complete(struct usb_ep *ep, struct
> usb_request *req)
>
> ?????? switch (req->status) {
> ?????? case 0:
> +?????? case -EXDEV: /* we ignore missed transfers */
Any time you see missed_isoc, you have data dropped. May want to check
to see what's going on.
> break;
>
> ?????? case -ESHUTDOWN:??????? /* disconnect from host. */
>
>
BR,
Thinh
Hi,
On Sat, Apr 11, 2020 at 12:59:47AM +0000, Thinh Nguyen wrote:
>Hi,
>
>Michael Grzeschik wrote:
>> On Fri, Apr 10, 2020 at 01:09:23AM +0000, Thinh Nguyen wrote:
>>> Hi,
>>>
>>> Michael Grzeschik wrote:
>>>> On Thu, Nov 14, 2019 at 08:11:56PM +0000, Thinh Nguyen wrote:
>>>>> Michael Olbrich wrote:
>>>>>> On Wed, Nov 13, 2019 at 07:14:59PM +0000, Thinh Nguyen wrote:
>>>>>>> Alan Stern wrote:
>>>>>>>> On Wed, 13 Nov 2019, Michael Olbrich wrote:
>>>>>>>>> On Wed, Nov 13, 2019 at 03:55:01AM +0000, Thinh Nguyen wrote:
>>>>>>>>>> Michael Olbrich wrote:
>>>>>>>>>>> Currently, most gadget drivers handle isoc transfers on a best
>>>>>>>>>>> effort
>>>>>>>>>>> bases: If the request queue runs empty, then there will simply
>>>>>>>>>>> be gaps in
>>>>>>>>>>> the isoc data stream.
>>>>>>>>>>>
>>>>>>>>>>> The UVC gadget depends on this behaviour. It simply provides
>>>>>>>>>>> new requests
>>>>>>>>>>> when video frames are available and assumes that they are sent
>>>>>>>>>>> as soon as
>>>>>>>>>>> possible.
>>>>>>>>>>>
>>>>>>>>>>> The dwc3 gadget currently works differently: It assumes that
>>>>>>>>>>> there is a
>>>>>>>>>>> contiguous stream of requests without any gaps. If a request is
>>>>>>>>>>> too late,
>>>>>>>>>>> then it is dropped by the hardware.
>>>>>>>>>>> For the UVC gadget this means that a live stream stops after
>>>>>>>>>>> the first
>>>>>>>>>>> frame because all following requests are late.
>>>>>>>>>> Can you explain little more how UVC gadget fails?
>>>>>>>>>> dwc3 controller expects a steady stream of data otherwise it
>>>>>>>>>> will result
>>>>>>>>>> in missed_isoc status, and it should be fine as long as new
>>>>>>>>>> requests are
>>>>>>>>>> queued. The controller doesn't just drop the request unless
>>>>>>>>>> there's some
>>>>>>>>>> other failure.
>>>>>>>>> UVC (with a live stream) does not fill the complete bandwidth
>>>>>>>>> of an
>>>>>>>>> isochronous endpoint. Let's assume for the example that one video
>>>>>>>>> frame
>>>>>>>>> fills 3 requests. Because it is a live stream, there will be a
>>>>>>>>> gap between
>>>>>>>>> video frames. This is unavoidable, especially for compressed
>>>>>>>>> video. So the
>>>>>>>>> UVC gadget will have requests for the frame numbers 1 2 3 5 6 7 9
>>>>>>>>> 10 11 13 14
>>>>>>>>> 15 and so on.
>>>>>>>>> The dwc3 hardware tries to send those with frame numbers 1 2 3 4
>>>>>>>>> 5 6 7 8 9
>>>>>>>>> 10 11 12. So except for the fist few requests, all are late and
>>>>>>>>> result in a
>>>>>>>>> missed_isoc. I tried to just ignore the missed_isoc but that did
>>>>>>>>> not work
>>>>>>>>> for me. I only received the first frame at the other end.
>>>>>>>>> Maybe I missing something here, i don't have access to the
>>>>>>>>> hardware
>>>>>>>>> documentation, so I can only guess from the existing driver.
>>>>>>> The reason I asked is because your patch doesn't seem to address the
>>>>>>> actual issue.
>>>>>>>
>>>>>>> For the 2 checks you do here
>>>>>>> 1. There are currently no requests queued in the hardware
>>>>>>> 2. The current frame number provided by DSTS does not match the
>>>>>>> frame
>>>>>>> ? ??? number returned by the last transfer.
>>>>>>>
>>>>>>> For #1, it's already done in the dwc3 driver. (check
>>>>>>> dwc3_gadget_endpoint_transfer_in_progress())
>>>>>> But that's only after a isoc_missed occurred. What exactly does that
>>>>>> mean?
>>>>>> Was the request transferred or not? My tests suggest that it was not
>>>>>> transferred, so I wanted to catch this before it happens.
>>>>>
>>>>> Missed_isoc status means that the controller did not move all the data
>>>>> in an interval.
>>>>
>>>> I read in some Processor documentation that in case the host tries to
>>>> fetch data from the client and no active TRB (HWO=1) is available the
>>>> XferInProgress Interrupt will be produced, with the missed status set.
>>>> This is done because the hardware will produce zero length packets
>>>> on its own, to keep the stream running.
>>>
>>> The controller only generates XferInProgress if it had processed a TRB
>>> (with specific control bits). For IN direction, if the controller is
>>> starved of TRB, it will send a ZLP if the host requests for data.
>>
>> Which control bits are those? ISOC-First, Chain and Last bits?
>>
>> I see the Scatter-Gather preparation is using these pattern.
>
>The IOC bit. You can check the programming guide for more detail on how
>to setup the TRBs, but what we have in dwc3 is fine.
OK.
>>>>>>> For #2, it's unlikely that DSTS current frame number will match
>>>>>>> with the
>>>>>>> XferNotReady's frame number. So this check doesn't mean much.
>>>>>> The frame number is also updated for each "Transfer In Progress"
>>>>>> interrupt.
>>>>>> If they match, then there a new request can still be queued
>>>>>> successfully.
>>>>>> Without this I got unnecessary stop/start transfers in the middle
>>>>>> of a
>>>>>> video frame. But maybe something else was wrong here. I'd need to
>>>>>> recheck.
>>>>>
>>>>> The reason they may not match is 1) the frame_number is only updated
>>>>> after the software handles the XferInProgress interrupt. Depends on
>>>>> system latency, that value may not be updated at the time that we
>>>>> check
>>>>> the frame_number.
>>>>> 2) This check doesn't work if the service interval is greater than 1
>>>>> uframe. That is, it doesn't have to match exactly the time to be
>>>>> consider not late. Though, the second reason can easily be fixed.
>>>>
>>>> In the empty trb case, after the Hardware has send enough zero packets
>>>> this
>>>> active transfer has to be stopped with endtransfer cmd. Because every
>>>> next
>>>> update transfer on that active transfer will likely lead to further
>>>> missed
>>>> transfers, as the newly updated trb will be handled to late anyway.
>>>
>>> The controller is expecting the function driver to feed TRBs to the
>>> controller for every interval. If it's late, then the controller will
>>> consider that data "missed_isoc".
>>>
>>> In your case, the UVC driver seems to queue requests to the controller
>>> driver as if it is bulk requests, and the UVC expects those data to go
>>> out at the time it queues. To achieve what UVC needs, then you may want
>>> to issue END_TRANSFER command before the next burst of data. This way,
>>> the controller can restart the isoc endpoint and not consider the next
>>> video frame data late. There are some corner cases that you need to
>>> watch out for. If you're going for this route, we can look further.
>>
>> Right, for now the drivers is doing:
>>
>> - Strart Transfer (with the right frame number) once.
>>
>> - Update Transfer On every ep_queue with the corresponding TRB
>> ?setting CHN = 0, IOC = 1, First-ISOC = 1
>>
>> - End Transfer is somehow not handled right for this case.
>>
>> See my first comment. I think dwc3_prepare_one_trb_sg does the proper
>> chain
>> handling already.
>>
>>> Also, you'd need provide a way for the UVC to communicate to the dwc3 to
>>> let it know to expect the next burst of data.
>>
>> Can the UVC not just enqueue one big sg-request, which will represent
>> one burst
>> and not one TRB. Also that is? what the SG code already seem to handle
>> right.
>
>Do you need SG? What size does video class driver setup its isoc URB? If
>it's superspeed, max is 48K, and depending on the type of platform
>you're running UVC on, you can setup a single 48K buffer per request if
>you want to do that. However, it's probably not a good idea since many
>host controllers can't even handle even close to 48K/uframe.
We wan't to transfer uncached 4K Video frames via UVC. So Scatter-Gather
is a must anyway.
>What I was saying is if UVC knows when to wait for the next video data,
>it can tell dwc3 to stop the isoc endpoint before queuing the next video
>data in a set of requests. If UVC doesn't know that, then it needs to
>tell dwc3 to change its handling of isoc requests.
In function/gadget/uvc_video.c we take a buffer and pump it into many
available requests as we find. On the way the driver can drain that
video frame queue, in that case we could stop the transfer. I have
prepared a patch where we have only one source, queueing new requests,
so this could work.
For now to limit the interrupt load on the dwc3 driver we already set
the request no_interrupt on every nth request, and make sure the
interrupt is on the very last one of the frame. But with that we still
sometimes run into missed transfers as the queued trb list somehow ends
up with a TRB not having the IOC bit set.
>>>> The odd thing here is, that I don't see the refered XferInProgress
>>>> Interrupts with the missed event, when the started_list is empty.
>>>
>>> See my first comment.
>>>
>>>>
>>>> But this would be the only case to fall into this condition and
>>>> handle it
>>>> properly. Like alredy assumed in the following code:
>>>>
>>>> static void dwc3_gadget_endpoint_transfer_in_progress(struct dwc3_ep
>>>> *dep,
>>>> ???????????? const struct dwc3_event_depevt *event)
>>>> {
>>>> ...
>>>>
>>>> ?????? if (event->status & DEPEVT_STATUS_MISSED_ISOC) {
>>>> ?????????????? status = -EXDEV;
>>>>
>>>> ?????????????? if (list_empty(&dep->started_list))
>>>> ?????????????????????? stop = true;
>>>> ?????? }
>>>>
>>>> ...
>>>>
>>>> ?????? if (stop)
>>>> ?????????????? dwc3_stop_active_transfer(dep, true, true);
>>>> ...
>>>> }
>>>>
>>>> In fact I did sometimes see these XferInProgress Interrupts on empty
>>>> trb queue
>>>> after I stoped the tansfer when the started_list was empty right after
>>>> ep_cleanup_completed_requests has moved all trbs out of the queue.
>>>>
>>>> These Interrupts appeared right after the ENDTRANSFER cmd was send.
>>>> (But I
>>>> could no verify this every time)
>>>
>>> If END_TRANSFER command completes, then you should not see
>>> XferInProgress event. The next event should ber XferNotReady.
>>
>> Right. This also stops, after the Command Complete Interrupt arrives.
>>
>>>> Anyways in that case these Interrupts are not useful anymore, as I
>>>> already
>>>> implied the same stop, with ENDTRANSFER after we know that there are
>>>> no other
>>>> trbs in the chain.
>>>>
>>>
>>> Just curious, do you know if there's a reason for UVC to behave this
>>> way? Seems like what it's trying to do is more for bulk. Maybe it wants
>>> bandwidth priority perhaps?
>>
>> I don't know, probably it was more likely for USB 2.0 controllers to
>> be handled
>> this way.
>>
>> As mentioned the current uvc code is also working when we add this
>> changes.
>>
>> diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c
>> index ec357f64f319..a5dc44f2e9d8 100644
>> --- a/drivers/usb/dwc3/gadget.c
>> +++ b/drivers/usb/dwc3/gadget.c
>> @@ -2629,6 +2629,9 @@ static void
>> dwc3_gadget_endpoint_transfer_in_progress(struct dwc3_ep *dep,
>>
>> ?????? dwc3_gadget_ep_cleanup_completed_requests(dep, event, status);
>>
>> +?????? if (list_empty(&dep->started_list))
>> +?????????????? stop = true;
>> +
>
>You should check the pending list too and either drop them or prepare
>those requests (maybe too late). This happens when there's no available
>TRB but UVC queues more requests.
>Also, make sure this change only applies for isoc.
I will send the patches for that so we can discuss the right handling
for that in a separate thread.
>This may work. Just keep in mind that the timing parameter of the
>XferNotReady will be expired by the time the UVC queues the next isoc
>request. (Maybe that's why you tried to check for the current frame
>number via DSTS instead in your first patch?).
>With the new changes in Felipe testing/next branch, this may be ok.
>
>
>> ?????? if (stop)
>> ?????????????? dwc3_stop_active_transfer(dep, true, true);
>> ?????? else if (dwc3_gadget_ep_should_continue(dep))
>>
>> diff --git a/drivers/usb/gadget/function/uvc_video.c
>> b/drivers/usb/gadget/function/uvc_video.c
>> index da6ba8ba4bca..a3dac5d91aae 100644
>> --- a/drivers/usb/gadget/function/uvc_video.c
>> +++ b/drivers/usb/gadget/function/uvc_video.c
>> @@ -183,6 +183,7 @@ uvc_video_complete(struct usb_ep *ep, struct
>> usb_request *req)
>>
>> ?????? switch (req->status) {
>> ?????? case 0:
>> +?????? case -EXDEV: /* we ignore missed transfers */
>
>Any time you see missed_isoc, you have data dropped. May want to check
>to see what's going on.
See my comment above.
>> break;
>>
>> ?????? case -ESHUTDOWN:??????? /* disconnect from host. */
>>
>>
>
>BR,
>Thinh
>
Regards,
Michael
--
Pengutronix e.K. | |
Steuerwalder Str. 21 | http://www.pengutronix.de/ |
31137 Hildesheim, Germany | Phone: +49-5121-206917-0 |
Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
Hi,
Michael Grzeschik wrote:
> Hi,
>
> On Sat, Apr 11, 2020 at 12:59:47AM +0000, Thinh Nguyen wrote:
>> Hi,
>>
>> Michael Grzeschik wrote:
>>> On Fri, Apr 10, 2020 at 01:09:23AM +0000, Thinh Nguyen wrote:
>>>> Hi,
>>>>
>>>> Michael Grzeschik wrote:
>>>>> On Thu, Nov 14, 2019 at 08:11:56PM +0000, Thinh Nguyen wrote:
>>>>>> Michael Olbrich wrote:
>>>>>>> On Wed, Nov 13, 2019 at 07:14:59PM +0000, Thinh Nguyen wrote:
>>>>>>>> Alan Stern wrote:
>>>>>>>>> On Wed, 13 Nov 2019, Michael Olbrich wrote:
>>>>>>>>>> On Wed, Nov 13, 2019 at 03:55:01AM +0000, Thinh Nguyen wrote:
>>>>>>>>>>> Michael Olbrich wrote:
>>>>>>>>>>>> Currently, most gadget drivers handle isoc transfers on a best
>>>>>>>>>>>> effort
>>>>>>>>>>>> bases: If the request queue runs empty, then there will simply
>>>>>>>>>>>> be gaps in
>>>>>>>>>>>> the isoc data stream.
>>>>>>>>>>>>
>>>>>>>>>>>> The UVC gadget depends on this behaviour. It simply provides
>>>>>>>>>>>> new requests
>>>>>>>>>>>> when video frames are available and assumes that they are sent
>>>>>>>>>>>> as soon as
>>>>>>>>>>>> possible.
>>>>>>>>>>>>
>>>>>>>>>>>> The dwc3 gadget currently works differently: It assumes that
>>>>>>>>>>>> there is a
>>>>>>>>>>>> contiguous stream of requests without any gaps. If a
>>>>>>>>>>>> request is
>>>>>>>>>>>> too late,
>>>>>>>>>>>> then it is dropped by the hardware.
>>>>>>>>>>>> For the UVC gadget this means that a live stream stops after
>>>>>>>>>>>> the first
>>>>>>>>>>>> frame because all following requests are late.
>>>>>>>>>>> Can you explain little more how UVC gadget fails?
>>>>>>>>>>> dwc3 controller expects a steady stream of data otherwise it
>>>>>>>>>>> will result
>>>>>>>>>>> in missed_isoc status, and it should be fine as long as new
>>>>>>>>>>> requests are
>>>>>>>>>>> queued. The controller doesn't just drop the request unless
>>>>>>>>>>> there's some
>>>>>>>>>>> other failure.
>>>>>>>>>> UVC (with a live stream) does not fill the complete bandwidth
>>>>>>>>>> of an
>>>>>>>>>> isochronous endpoint. Let's assume for the example that one
>>>>>>>>>> video
>>>>>>>>>> frame
>>>>>>>>>> fills 3 requests. Because it is a live stream, there will be a
>>>>>>>>>> gap between
>>>>>>>>>> video frames. This is unavoidable, especially for compressed
>>>>>>>>>> video. So the
>>>>>>>>>> UVC gadget will have requests for the frame numbers 1 2 3 5 6
>>>>>>>>>> 7 9
>>>>>>>>>> 10 11 13 14
>>>>>>>>>> 15 and so on.
>>>>>>>>>> The dwc3 hardware tries to send those with frame numbers 1 2 3 4
>>>>>>>>>> 5 6 7 8 9
>>>>>>>>>> 10 11 12. So except for the fist few requests, all are late and
>>>>>>>>>> result in a
>>>>>>>>>> missed_isoc. I tried to just ignore the missed_isoc but that did
>>>>>>>>>> not work
>>>>>>>>>> for me. I only received the first frame at the other end.
>>>>>>>>>> Maybe I missing something here, i don't have access to the
>>>>>>>>>> hardware
>>>>>>>>>> documentation, so I can only guess from the existing driver.
>>>>>>>> The reason I asked is because your patch doesn't seem to
>>>>>>>> address the
>>>>>>>> actual issue.
>>>>>>>>
>>>>>>>> For the 2 checks you do here
>>>>>>>> 1. There are currently no requests queued in the hardware
>>>>>>>> 2. The current frame number provided by DSTS does not match the
>>>>>>>> frame
>>>>>>>> ? ??? number returned by the last transfer.
>>>>>>>>
>>>>>>>> For #1, it's already done in the dwc3 driver. (check
>>>>>>>> dwc3_gadget_endpoint_transfer_in_progress())
>>>>>>> But that's only after a isoc_missed occurred. What exactly does
>>>>>>> that
>>>>>>> mean?
>>>>>>> Was the request transferred or not? My tests suggest that it was
>>>>>>> not
>>>>>>> transferred, so I wanted to catch this before it happens.
>>>>>>
>>>>>> Missed_isoc status means that the controller did not move all the
>>>>>> data
>>>>>> in an interval.
>>>>>
>>>>> I read in some Processor documentation that in case the host tries to
>>>>> fetch data from the client and no active TRB (HWO=1) is available the
>>>>> XferInProgress Interrupt will be produced, with the missed status
>>>>> set.
>>>>> This is done because the hardware will produce zero length packets
>>>>> on its own, to keep the stream running.
>>>>
>>>> The controller only generates XferInProgress if it had processed a TRB
>>>> (with specific control bits). For IN direction, if the controller is
>>>> starved of TRB, it will send a ZLP if the host requests for data.
>>>
>>> Which control bits are those? ISOC-First, Chain and Last bits?
>>>
>>> I see the Scatter-Gather preparation is using these pattern.
>>
>> The IOC bit. You can check the programming guide for more detail on how
>> to setup the TRBs, but what we have in dwc3 is fine.
>
> OK.
>
>>>>>>>> For #2, it's unlikely that DSTS current frame number will match
>>>>>>>> with the
>>>>>>>> XferNotReady's frame number. So this check doesn't mean much.
>>>>>>> The frame number is also updated for each "Transfer In Progress"
>>>>>>> interrupt.
>>>>>>> If they match, then there a new request can still be queued
>>>>>>> successfully.
>>>>>>> Without this I got unnecessary stop/start transfers in the middle
>>>>>>> of a
>>>>>>> video frame. But maybe something else was wrong here. I'd need to
>>>>>>> recheck.
>>>>>>
>>>>>> The reason they may not match is 1) the frame_number is only updated
>>>>>> after the software handles the XferInProgress interrupt. Depends on
>>>>>> system latency, that value may not be updated at the time that we
>>>>>> check
>>>>>> the frame_number.
>>>>>> 2) This check doesn't work if the service interval is greater than 1
>>>>>> uframe. That is, it doesn't have to match exactly the time to be
>>>>>> consider not late. Though, the second reason can easily be fixed.
>>>>>
>>>>> In the empty trb case, after the Hardware has send enough zero
>>>>> packets
>>>>> this
>>>>> active transfer has to be stopped with endtransfer cmd. Because every
>>>>> next
>>>>> update transfer on that active transfer will likely lead to further
>>>>> missed
>>>>> transfers, as the newly updated trb will be handled to late anyway.
>>>>
>>>> The controller is expecting the function driver to feed TRBs to the
>>>> controller for every interval. If it's late, then the controller will
>>>> consider that data "missed_isoc".
>>>>
>>>> In your case, the UVC driver seems to queue requests to the controller
>>>> driver as if it is bulk requests, and the UVC expects those data to go
>>>> out at the time it queues. To achieve what UVC needs, then you may
>>>> want
>>>> to issue END_TRANSFER command before the next burst of data. This way,
>>>> the controller can restart the isoc endpoint and not consider the next
>>>> video frame data late. There are some corner cases that you need to
>>>> watch out for. If you're going for this route, we can look further.
>>>
>>> Right, for now the drivers is doing:
>>>
>>> - Strart Transfer (with the right frame number) once.
>>>
>>> - Update Transfer On every ep_queue with the corresponding TRB
>>> ?setting CHN = 0, IOC = 1, First-ISOC = 1
>>>
>>> - End Transfer is somehow not handled right for this case.
>>>
>>> See my first comment. I think dwc3_prepare_one_trb_sg does the proper
>>> chain
>>> handling already.
>>>
>>>> Also, you'd need provide a way for the UVC to communicate to the
>>>> dwc3 to
>>>> let it know to expect the next burst of data.
>>>
>>> Can the UVC not just enqueue one big sg-request, which will represent
>>> one burst
>>> and not one TRB. Also that is? what the SG code already seem to handle
>>> right.
>>
>> Do you need SG? What size does video class driver setup its isoc URB? If
>> it's superspeed, max is 48K, and depending on the type of platform
>> you're running UVC on, you can setup a single 48K buffer per request if
>> you want to do that. However, it's probably not a good idea since many
>> host controllers can't even handle even close to 48K/uframe.
>
> We wan't to transfer uncached 4K Video frames via UVC. So Scatter-Gather
> is a must anyway.
>
>> What I was saying is if UVC knows when to wait for the next video data,
>> it can tell dwc3 to stop the isoc endpoint before queuing the next video
>> data in a set of requests. If UVC doesn't know that, then it needs to
>> tell dwc3 to change its handling of isoc requests.
>
> In function/gadget/uvc_video.c we take a buffer and pump it into many
> available requests as we find. On the way the driver can drain that
> video frame queue, in that case we could stop the transfer. I have
> prepared a patch where we have only one source, queueing new requests,
> so this could work.
>
> For now to limit the interrupt load on the dwc3 driver we already set
> the request no_interrupt on every nth request, and make sure the
> interrupt is on the very last one of the frame. But with that we still
> sometimes run into missed transfers as the queued trb list somehow ends
> up with a TRB not having the IOC bit set.
>
>>>>> The odd thing here is, that I don't see the refered XferInProgress
>>>>> Interrupts with the missed event, when the started_list is empty.
>>>>
>>>> See my first comment.
>>>>
>>>>>
>>>>> But this would be the only case to fall into this condition and
>>>>> handle it
>>>>> properly. Like alredy assumed in the following code:
>>>>>
>>>>> static void dwc3_gadget_endpoint_transfer_in_progress(struct dwc3_ep
>>>>> *dep,
>>>>> ???????????? const struct dwc3_event_depevt *event)
>>>>> {
>>>>> ...
>>>>>
>>>>> ?????? if (event->status & DEPEVT_STATUS_MISSED_ISOC) {
>>>>> ?????????????? status = -EXDEV;
>>>>>
>>>>> ?????????????? if (list_empty(&dep->started_list))
>>>>> ?????????????????????? stop = true;
>>>>> ?????? }
>>>>>
>>>>> ...
>>>>>
>>>>> ?????? if (stop)
>>>>> ?????????????? dwc3_stop_active_transfer(dep, true, true);
>>>>> ...
>>>>> }
>>>>>
>>>>> In fact I did sometimes see these XferInProgress Interrupts on empty
>>>>> trb queue
>>>>> after I stoped the tansfer when the started_list was empty right
>>>>> after
>>>>> ep_cleanup_completed_requests has moved all trbs out of the queue.
>>>>>
>>>>> These Interrupts appeared right after the ENDTRANSFER cmd was send.
>>>>> (But I
>>>>> could no verify this every time)
>>>>
>>>> If END_TRANSFER command completes, then you should not see
>>>> XferInProgress event. The next event should ber XferNotReady.
>>>
>>> Right. This also stops, after the Command Complete Interrupt arrives.
>>>
>>>>> Anyways in that case these Interrupts are not useful anymore, as I
>>>>> already
>>>>> implied the same stop, with ENDTRANSFER after we know that there are
>>>>> no other
>>>>> trbs in the chain.
>>>>>
>>>>
>>>> Just curious, do you know if there's a reason for UVC to behave this
>>>> way? Seems like what it's trying to do is more for bulk. Maybe it
>>>> wants
>>>> bandwidth priority perhaps?
>>>
>>> I don't know, probably it was more likely for USB 2.0 controllers to
>>> be handled
>>> this way.
>>>
>>> As mentioned the current uvc code is also working when we add this
>>> changes.
>>>
>>> diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c
>>> index ec357f64f319..a5dc44f2e9d8 100644
>>> --- a/drivers/usb/dwc3/gadget.c
>>> +++ b/drivers/usb/dwc3/gadget.c
>>> @@ -2629,6 +2629,9 @@ static void
>>> dwc3_gadget_endpoint_transfer_in_progress(struct dwc3_ep *dep,
>>>
>>> ?????? dwc3_gadget_ep_cleanup_completed_requests(dep, event, status);
>>>
>>> +?????? if (list_empty(&dep->started_list))
>>> +?????????????? stop = true;
>>> +
>>
>> You should check the pending list too and either drop them or prepare
>> those requests (maybe too late). This happens when there's no available
>> TRB but UVC queues more requests.
>> Also, make sure this change only applies for isoc.
>
> I will send the patches for that so we can discuss the right handling
> for that in a separate thread.
Sure, we can review further then.
BR,
Thinh