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`?
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.
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.
Thomas
Regards,
Hans
+
+ 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);
