Received-SPF: pass (google.com: domain of linux-kernel+bounces-211190-linux.lists.archive=gmail.com@vger.kernel.org designates 139.178.88.99 as permitted sender) client-ip=139.178.88.99;
Precedence: bulk
MIME-Version: 1.0
References: <20240323-resend-hwtimestamp-v10-0-b08e590d97c7@chromium.org>
 <20240323-resend-hwtimestamp-v10-4-b08e590d97c7@chromium.org>
 <4kck7oedsnj6kfiv7ykwsjg35qodg5bdktu5t5w3xtg2xuscto@2yh6kfdqwimc>
 <20240610114306.GR18479@pendragon.ideasonboard.com> <CAAFQd5DAXq6fTrp6jF42URrwzwE+tGz_jJCRM2bhieD76u+QpA@mail.gmail.com>
 <20240612074342.GA28989@pendragon.ideasonboard.com>
In-Reply-To: <20240612074342.GA28989@pendragon.ideasonboard.com>
From: Tomasz Figa <tfiga@chromium.org>
Date: Wed, 12 Jun 2024 17:35:38 +0900
Message-ID: <CAAFQd5AS_h_ydkNnKU-y--Cd_SSoMyyQhj6tXeZ_oydoqzwFFw@mail.gmail.com>
Subject: Re: [PATCH v10 4/6] media: uvcvideo: Allow hw clock updates with
 buffers not full
To: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Cc: Ricardo Ribalda <ribalda@chromium.org>, Mauro Carvalho Chehab <mchehab@kernel.org>, 
	linux-media@vger.kernel.org, linux-kernel@vger.kernel.org, 
	"hn.chen" <hn.chen@sunplusit.com>, Hans Verkuil <hverkuil@xs4all.nl>, 
	Sergey Senozhatsky <senozhatsky@chromium.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Wed, Jun 12, 2024 at 4:44=E2=80=AFPM Laurent Pinchart
<laurent.pinchart@ideasonboard.com> wrote:
>
> On Wed, Jun 12, 2024 at 12:28:56PM +0900, Tomasz Figa wrote:
> > On Mon, Jun 10, 2024 at 8:43=E2=80=AFPM Laurent Pinchart wrote:
> > > On Wed, May 29, 2024 at 05:03:08PM +0900, Tomasz Figa wrote:
> > > > On Sat, Mar 23, 2024 at 10:48:05AM +0000, Ricardo Ribalda wrote:
> > > > > With UVC 1.5 we get as little as one clock sample per frame. Whic=
h means
> > > > > that it takes 32 frames to move from the software timestamp to th=
e
> > > > > hardware timestamp method.
> > > > >
> > > > > This results in abrupt changes in the timestamping after 32 frame=
s (~1
> > > > > second), resulting in noticeable artifacts when used for encoding=
.
> > > > >
> > > > > With this patch we modify the update algorithm to work with whate=
ver
> > > > > amount of values are available.
> > > > >
> > > > > Tested-by: HungNien Chen <hn.chen@sunplusit.com>
> > > > > Reviewed-by: Sergey Senozhatsky <senozhatsky@chromium.org>
> > > > > Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
> > > > > Signed-off-by: Ricardo Ribalda <ribalda@chromium.org>
> > > > > ---
> > > > >  drivers/media/usb/uvc/uvc_video.c | 16 ++++++++++++++--
> > > > >  1 file changed, 14 insertions(+), 2 deletions(-)
> > > > >
> > > > > diff --git a/drivers/media/usb/uvc/uvc_video.c b/drivers/media/us=
b/uvc/uvc_video.c
> > > > > index d6ca383f643e3..af25b9f1b53fe 100644
> > > > > --- a/drivers/media/usb/uvc/uvc_video.c
> > > > > +++ b/drivers/media/usb/uvc/uvc_video.c
> > > > > @@ -768,10 +768,10 @@ void uvc_video_clock_update(struct uvc_stre=
aming *stream,
> > > > >
> > > > >     spin_lock_irqsave(&clock->lock, flags);
> > > > >
> > > > > -   if (clock->count < clock->size)
> > > > > +   if (clock->count < 2)
> > > > >             goto done;
> > > > >
> > > > > -   first =3D &clock->samples[clock->head];
> > > > > +   first =3D &clock->samples[(clock->head - clock->count + clock=
->size) % clock->size];
> > > > >     last =3D &clock->samples[(clock->head - 1 + clock->size) % cl=
ock->size];
> > > > >
> > > > >     /* First step, PTS to SOF conversion. */
> > > > > @@ -786,6 +786,18 @@ void uvc_video_clock_update(struct uvc_strea=
ming *stream,
> > > > >     if (y2 < y1)
> > > > >             y2 +=3D 2048 << 16;
> > > > >
> > > > > +   /*
> > > > > +    * Have at least 1/4 of a second of timestamps before we
> > > > > +    * try to do any calculation. Otherwise we do not have enough
> > > > > +    * precision. This value was determined by running Android CT=
S
> > > > > +    * on different devices.
> > > > > +    *
> > > > > +    * dev_sof runs at 1KHz, and we have a fixed point precision =
of
> > > > > +    * 16 bits.
> > > > > +    */
> > > > > +   if ((y2 - y1) < ((1000 / 4) << 16))
> > > > > +           goto done;
> > > >
> > > > Not a comment for this patch directly, but...
> > > >
> > > > This kind of makes me wonder if we don't want to have some document=
ation
> > > > that specifies what the userspace can expect from the timestamps, s=
o
> > > > that this isn't changed randomly in the future breaking what was fi=
xed
> > > > by this patch.
> > >
> > > I think timestamp handling should really be moved to userspace. It wi=
ll
> > > be easier to handle with floating-point arithmetic there. That would
> > > have been difficult to manage for applications a while ago, but now t=
hat
> > > we have libcamera, we could implement it there. This isn't high on my
> > > todo list though.
> >
> > While indeed that would probably be a better way to handle the complex
> > logic if we designed the driver today, we already have userspace that
> > expects the timestamps to be handled correctly in the kernel. I
> > suspect moving it to the userspace would require some core V4L2
> > changes to define a new timestamp handling mode, where multiple raw
> > hardware timestamps are exposed to the userspace, instead of the high
> > level system monotonic one.
>
> The uvcvideo driver already supports exposing the packet headers to
> userspace through a metadata capture device, so I think we have all the
> components we need. The only missing thing would be the implementation
> in libcamera :-)

Okay, I see. That would make it easier indeed. :)

That said, we still need to provide some valid timestamps in the
v4l2_buffer struct returned from DQBUF, as per the current API
contract, so we can't simply remove the timestamp handling code from
the kernel completely.