On 16/11/2020 16:31, Laurent Pinchart wrote:
> Hello everybody,
>
> Sorry for chiming in late.
>
> On Tue, Nov 10, 2020 at 05:01:31PM +0100, Hans Verkuil wrote:
>> On 10/11/2020 16:50, Thomas Hämmerle wrote:
>>> On 10.11.20 16:36, Hans Verkuil wrote:
>>>> On 10/11/2020 16:10, Thomas Hämmerle wrote:
>>>>> On 10.11.20 15:43, Hans Verkuil wrote:
>>>>>> On 10/11/2020 12:53, Thomas Hämmerle wrote:
>>>>>>> On 10.11.20 11:31, Hans Verkuil wrote:
>>>>>>>> On 10/11/2020 09:25, thomas.haemmerle@xxxxxxxxxxxxxx wrote:
>>>>>>>>> From: Thomas Haemmerle <thomas.haemmerle@xxxxxxxxxxxxxx>
>>>>>>>>>
>>>>>>>>> Currently, the UVC function is activated when open on the corresponding
>>>>>>>>> v4l2 device is called.
>>>>>>>>> On another open the activation of the function fails since the
>>>>>>>>> deactivation counter in `usb_function_activate` equals 0. However the
>>>>>>>>> error is not returned to userspace since the open of the v4l2 device is
>>>>>>>>> successful.
>>>>>>>>>
>>>>>>>>> On a close the function is deactivated (since deactivation counter still
>>>>>>>>> equals 0) and the video is disabled in `uvc_v4l2_release`, although
>>>>>>>>> another process potentially is streaming.
>>>>>>>>>
>>>>>>>>> Move activation of UVC function to subscription on UVC_EVENT_SETUP and
>>>>>>>>> keep track of the number of subscribers (limited to 1) because there we
>>>>>>>>> can guarantee for a userspace program utilizing UVC.
>>>>>>>>> Extend the `struct uvc_file_handle` with member `bool connected` to tag
>>>>>>>>> it for a deactivation of the function.
>>>>>>>>>
>>>>>>>>> With this a process is able to check capabilities of the v4l2 device
>>>>>>>>> without deactivating the function for another process actually using the
>>>>>>>>> device for UVC streaming.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Thomas Haemmerle <thomas.haemmerle@xxxxxxxxxxxxxx>
>>>>>>>>> ---
>>>>>>>>> v2:
>>>>>>>>> - fix deadlock in `uvc_v4l2_unsubscribe_event()` (mutex is already
>>>>>>>>> locked in v4l2-core) introduced in v1
>>>>>>>>> - lock mutex in `uvc_v4l2_release()` to suppress ioctls and protect
>>>>>>>>> connected
>>>>>>>>>
>>>>>>>>> drivers/usb/gadget/function/uvc.h | 2 +
>>>>>>>>> drivers/usb/gadget/function/uvc_v4l2.c | 56 +++++++++++++++++++++-----
>>>>>>>>> 2 files changed, 48 insertions(+), 10 deletions(-)
>>>>>>>>>
>>>>>>>>> diff --git a/drivers/usb/gadget/function/uvc.h b/drivers/usb/gadget/function/uvc.h
>>>>>>>>> index 73da4f9a8d4c..0d0bcbffc8fd 100644
>>>>>>>>> --- a/drivers/usb/gadget/function/uvc.h
>>>>>>>>> +++ b/drivers/usb/gadget/function/uvc.h
>>>>>>>>> @@ -117,6 +117,7 @@ struct uvc_device {
>>>>>>>>> enum uvc_state state;
>>>>>>>>> struct usb_function func;
>>>>>>>>> struct uvc_video video;
>>>>>>>>> + unsigned int connections;
>>>>>>>>>
>>>>>>>>> /* Descriptors */
>>>>>>>>> struct {
>>>>>>>>> @@ -147,6 +148,7 @@ static inline struct uvc_device *to_uvc(struct usb_function *f)
>>>>>>>>> struct uvc_file_handle {
>>>>>>>>> struct v4l2_fh vfh;
>>>>>>>>> struct uvc_video *device;
>>>>>>>>> + bool connected;
>>>>>>>>> };
>>>>>>>>>
>>>>>>>>> #define to_uvc_file_handle(handle) \
>>>>>>>>> diff --git a/drivers/usb/gadget/function/uvc_v4l2.c b/drivers/usb/gadget/function/uvc_v4l2.c
>>>>>>>>> index 67922b1355e6..aee4888e17b1 100644
>>>>>>>>> --- a/drivers/usb/gadget/function/uvc_v4l2.c
>>>>>>>>> +++ b/drivers/usb/gadget/function/uvc_v4l2.c
>>>>>>>>> @@ -228,17 +228,57 @@ static int
>>>>>>>>> uvc_v4l2_subscribe_event(struct v4l2_fh *fh,
>>>>>>>>> const struct v4l2_event_subscription *sub)
>>>>>>>>> {
>>>>>>>>> + struct uvc_device *uvc = video_get_drvdata(fh->vdev);
>>>>>>>>> + struct uvc_file_handle *handle = to_uvc_file_handle(fh);
>>>>>>>>> + int ret;
>>>>>>>>> +
>>>>>>>>> if (sub->type < UVC_EVENT_FIRST || sub->type > UVC_EVENT_LAST)
>>>>>>>>> return -EINVAL;
>>>>>>>>>
>>>>>>>>> - return v4l2_event_subscribe(fh, sub, 2, NULL);
>>>>>>>>> + if ((sub->type == UVC_EVENT_SETUP) && (uvc->connections >= 1))
>>>>>>>>> + return -EBUSY;
>>>>>>>>> +
>>>>>>>>> + ret = v4l2_event_subscribe(fh, sub, 2, NULL);
>>>>>>>>> + if (ret < 0)
>>>>>>>>> + return ret;
>>>>>>>>> +
>>>>>>>>> + if (sub->type == UVC_EVENT_SETUP) {
>>>>>>>>> + uvc->connections++;
>>>>>>>>> + handle->connected = true;
>>>>>>>>> + uvc_function_connect(uvc);
>>>>>>>>> + }
>>>>>>>>
>>>>>>>> This makes no sense. Why would subscribing to a SETUP event
>>>>>>>> mean that you are 'connected'?
>>>>>>>
>>>>>>> Subscribing to a SETUP does not mean that you are connected - on a
>>>>>>> subscription to SETUP we can expect that there is a userspace process,
>>>>>>> utilizing UVC -- which is required for the UVC gadget function -- and
>>>>>>> everything is ready to enable the function by calling uvc_function_connect.
>>>>>>> How about calling it `function_connected`?
>>>>>>
>>>>>> Any application can open the device node and subscribe to this event.
>>>>>> This is not a good place to call uvc_function_connect(), it's an abuse of
>>>>>> the V4L2 API.
>>>>>
>>>>> Of course - any application can subscribe to this event but since the
>>>>> event is defined in g_uvc.h I thought that a subscription is most likely
>>>>> done by a UVC application.
>
> I agree with Thomas here, subscribing to UVC_EVENT_SETUP means that the
> application is a UVC gadget application. It's a bit of a hack, but short
> of introducing a custom ioctl, I think it's a fairly good option.
Yuck.
OK, if you do this, then this should be documented, both in the driver and
esp. in the public g_uvc.h header. This is completely non-standard and
unexpected behavior.
It would be nice BTW if all these UVC events are documented in the header.
Regards,
Hans
>
>>>>>> I dug a bit more into this (I have very little gadget driver experience),
>>>>>> and isn't calling uvc_function_connect() something that should be done
>>>>>> on the first open? And when the last filehandle is closed, then
>>>>>> uvc_function_disconnect() can be called to clean up?
>>>>>
>>>>> Isn't the possibility of the delayed activation of the usb function in
>>>>> composite.c intended to wait for a corresponding userspace application,
>>>>> which is required to make the gadget work?
>>>>>
>>>>> If so, a simple open of the v4l2 device does not mean that the gadget is
>>>>> ready to go: starting pipewire for example results in quering the device
>>>>> capabilities of all devices. But this does not mean that the gadget will
>>>>> work.
>>>>> Therefore I was looking for a place, where we can expect, that the
>>>>> application opened the device will utilize UVC and found that a
>>>>> subscription to a UVC specific event is the right place.
>>>>
>>>> That's why I suggested to do this when buffers are allocated. That's when
>>>> the application commits resources to actually use the device. Until that
>>>> point everything that an application does is just querying or configuring,
>>>> but not actually using it.
>>>
>>> The userspace application implementations (including the one from
>>> Laurent Pinchart) usually request buffers on the UVC events sent by the
>>> host. Therefore we would run into chicken-and-egg problem.
>>
>> Ah, stupid of me. You're right.
>>
>> For now I go with calling uvc_function_connect() on the first open. If
>> nothing else, this is compatible with the current situation, while still
>> allowing for multiple opens.
>
> The problem with connecting on first open is that any application
> querying the device (v4l_id from udev, pipewire as Thomas mentioned,
> ...) will generate spurious connect/disconnect events.
>
> This being said, the current code already suffers from the spurious
> connect/disconnect issue, so we could possibly fix the race condition
> first with a patch that calls uvc_function_connect() on the first open
> and uvc_function_disconnect() on last close, and then move the logic to
> event subscription in a separate patch. Not sure it would be worth it
> though.
>
>>>> BTW, Laurent Pinchart who maintains this driver is away for a week, so he
>>>> might know more about this when he comes back.
>>>
>>> Okay, so I suggest to wait for him.
>>>
>>> Thanks for the review so far!
>>>
>>>>>> That would make much more sense. Grep for v4l2_fh_is_singular_file() on how
>>>>>> to do this, quite a few media drivers need this.
>>>>>>
>>>>>>>> It should be possible to open a V4L2 device node any number of times,
>>>>>>>> and any filehandle can subscribe to any event, but typically once
>>>>>>>> userspace allocates buffers (VIDIOC_REQBUFS or VIDIOC_CREATE_BUFS)
>>>>>>>> then that filehandle is marked as the owner of the device and other
>>>>>>>> open filehandles are no longer allowed to allocate buffers or stream video.
>>>>>>>
>>>>>>> Sure - this can be also done if userspace allocates buffers.
>>>>>>> But why does it make more sense to call uvc_function_connect on
>>>>>>> VIDIOC_REQBUFS or VIDIOC_CREATE_BUFS instead of subscribtion to a UVC event?
>>>>>>>
>>>>>>>> See e.g. drivers/media/common/videobuf2/videobuf2-v4l2.c
>>>>>>>> and vb2_ioctl_reqbufs and other vb2_ioctl_* functions.
>>>>>>>>
>>>>>>>> Unfortunately this UVC gadget driver is rather old and is not using
>>>>>>>> these helper functions.
>>>>>>>>
>>>>>>>> Running 'v4l2-compliance' will likely fail on a lot of tests for this
>>>>>>>> driver.
>
> I don't think this driver will ever be able to pass v4l2-compliance as
> it's not a typical V4L2 output device. It uses V4L2 as that was the best
> match for a userspace API, but custom ioctls and events are required to
> drive the device, and the userspace application needs to work
> hand-in-hand with the driver to respond to UVC requests originating from
> the host. No other V4L2 application will be able to use the device.
>
> This being said, the driver is indeed old, and I'm sure improvements are
> possible, but full v4l2-compliance isn't a reasonable goal.
>
>>>>>>>> This driver probably could use some TLC.
>>>>>>>
>>>>>>> I totally agree with that, however this change fixes at least one test
>>>>>>> of 'v4l2-compliance'.
>>>>>>> Currently a running UVC connection can be corrupted by another process
>>>>>>> which just opens and closes the device.
>>>>>>>
>>>>>>>>> +
>>>>>>>>> + return 0;
>>>>>>>>> +}
>>>>>>>>> +
>>>>>>>>> +static void uvc_v4l2_disable(struct uvc_device *uvc)
>>>>>>>>> +{
>>>>>>>>> + if (--uvc->connections)
>>>>>>>>> + return;
>>>>>>>>> +
>>>>>>>>> + uvc_function_disconnect(uvc);
>>>>>>>>> + uvcg_video_enable(&uvc->video, 0);
>>>>>>>>> + uvcg_free_buffers(&uvc->video.queue);
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>> static int
>>>>>>>>> uvc_v4l2_unsubscribe_event(struct v4l2_fh *fh,
>>>>>>>>> const struct v4l2_event_subscription *sub)
>>>>>>>>> {
>>>>>>>>> - return v4l2_event_unsubscribe(fh, sub);
>>>>>>>>> + struct uvc_device *uvc = video_get_drvdata(fh->vdev);
>>>>>>>>> + struct uvc_file_handle *handle = to_uvc_file_handle(fh);
>>>>>>>>> + int ret;
>>>>>>>>> +
>>>>>>>>> + ret = v4l2_event_unsubscribe(fh, sub);
>>>>>>>>> + if (ret < 0)
>>>>>>>>> + return ret;
>>>>>>>>> +
>>>>>>>>> + if ((sub->type == UVC_EVENT_SETUP) && handle->connected) {
>>>>>>>>> + uvc_v4l2_disable(uvc);
>>>>>>>>> + handle->connected = false;
>>>>>>>>> + }
>>>>>>>>> +
>>>>>>>>> + return 0;
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>> static long
>>>>>>>>> @@ -293,7 +333,6 @@ uvc_v4l2_open(struct file *file)
>>>>>>>>> handle->device = &uvc->video;
>>>>>>>>> file->private_data = &handle->vfh;
>>>>>>>>>
>>>>>>>>> - uvc_function_connect(uvc);
>>>>>>>>> return 0;
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>> @@ -303,14 +342,11 @@ uvc_v4l2_release(struct file *file)
>>>>>>>>> struct video_device *vdev = video_devdata(file);
>>>>>>>>> struct uvc_device *uvc = video_get_drvdata(vdev);
>>>>>>>>> struct uvc_file_handle *handle = to_uvc_file_handle(file->private_data);
>>>>>>>>> - struct uvc_video *video = handle->device;
>>>>>>>>> -
>>>>>>>>> - uvc_function_disconnect(uvc);
>>>>>>>>>
>>>>>>>>> - mutex_lock(&video->mutex);
>>>>>>>>> - uvcg_video_enable(video, 0);
>>>>>>>>> - uvcg_free_buffers(&video->queue);
>>>>>>>>> - mutex_unlock(&video->mutex);
>>>>>>>>> + mutex_lock(&uvc->video.mutex);
>>>>>>>>> + if (handle->connected)
>>>>>>>>> + uvc_v4l2_disable(uvc);
>>>>>>>>> + mutex_unlock(&uvc->video.mutex);
>>>>>>>>>
>>>>>>>>> file->private_data = NULL;
>>>>>>>>> v4l2_fh_del(&handle->vfh);
>
