Re: [PATCH] media: v4l2-mem2mem: always call poll_wait() on queues

[Alexandre Courbot]

On Thu, Nov 5, 2020 at 11:05 PM Alexandre Courbot <gnurou@xxxxxxxxx> wrote:
>
> On Thu, Nov 5, 2020 at 10:12 PM Hans Verkuil <hverkuil-cisco@xxxxxxxxx> wrote:
> >
> > On 05/11/2020 13:52, Alexandre Courbot wrote:
> > > On Thu, Nov 5, 2020 at 9:36 PM Hans Verkuil <hverkuil-cisco@xxxxxxxxx> wrote:
> > >>
> > >> On 05/11/2020 13:21, Alexandre Courbot wrote:
> > >>> On Tue, Nov 3, 2020 at 6:48 PM Hans Verkuil <hverkuil-cisco@xxxxxxxxx> wrote:
> > >>>>
> > >>>> On 03/11/2020 09:51, Alexandre Courbot wrote:
> > >>>>> Hi Hans,
> > >>>>>
> > >>>>> On Sat, Oct 31, 2020 at 12:09 AM Hans Verkuil <hverkuil-cisco@xxxxxxxxx> wrote:
> > >>>>>>
> > >>>>>> On 22/10/2020 14:24, Alexandre Courbot wrote:
> > >>>>>>> do_poll()/do_select() seem to set the _qproc member of poll_table to
> > >>>>>>> NULL the first time they are called on a given table, making subsequent
> > >>>>>>> calls of poll_wait() on that table no-ops. This is a problem for mem2mem
> > >>>>>>> which calls poll_wait() on the V4L2 queues' waitqueues only when a
> > >>>>>>> queue-related event is requested, which may not necessarily be the case
> > >>>>>>> during the first poll.
> > >>>>>>>
> > >>>>>>> For instance, a stateful decoder is typically only interested in
> > >>>>>>> EPOLLPRI events when it starts, and will switch to listening to both
> > >>>>>>> EPOLLPRI and EPOLLIN after receiving the initial resolution change event
> > >>>>>>> and configuring the CAPTURE queue. However by the time that switch
> > >>>>>>> happens and v4l2_m2m_poll_for_data() is called for the first time,
> > >>>>>>> poll_wait() has become a no-op and the V4L2 queues waitqueues thus
> > >>>>>>> cannot be registered.
> > >>>>>>>
> > >>>>>>> Fix this by moving the registration to v4l2_m2m_poll() and do it whether
> > >>>>>>> or not one of the queue-related events are requested.
> > >>>>>>
> > >>>>>> This looks good, but would it be possible to add a test for this to
> > >>>>>> v4l2-compliance? (Look for POLL_MODE_EPOLL in v4l2-test-buffers.cpp)
> > >>>>>>
> > >>>>>> If I understand this right, calling EPOLL_CTL_ADD for EPOLLPRI, then
> > >>>>>> calling EPOLL_CTL_ADD for EPOLLIN/OUT would trigger this? Or does there
> > >>>>>> have to be an epoll_wait call in between?
> > >>>>>
> > >>>>> Even without an epoll_wait() in between the behavior is visible.
> > >>>>> v4l2_m2m_poll() will be called once during the initial EPOLL_CTL_ADD
> > >>>>> and this will trigger the bug.
> > >>>>>
> > >>>>>> Another reason for adding this test is that I wonder if regular capture
> > >>>>>> or output V4L2 devices don't have the same issue.
> > >>>>>>
> > >>>>>> It's a very subtle bug and so adding a test for this to v4l2-compliance
> > >>>>>> would be very useful.
> > >>>>>
> > >>>>> I fully agree, this is very counter-intuitive since what basically
> > >>>>> happens is that the kernel's poll_wait() function becomes a no-op
> > >>>>> after the poll() hook of a driver is called for the first time. There
> > >>>>> is no way one can expect this behavior just from browsing the code so
> > >>>>> this is likely to affect other drivers.
> > >>>>>
> > >>>>> As for the test itself, we can easily reproduce the conditions for
> > >>>>> failure in v4l2-test-buffers.cpp's captureBufs() function, but doing
> > >>>>> so will make the streaming tests fail without being specific about the
> > >>>>> cause. Or maybe we should add another pollmode to specifically test
> > >>>>> epoll in this setup? Can I get your thoughts?
> > >>>>
> > >>>> No, just keep it as part of the poll test. Just add comments at the place
> > >>>> where it fails describing this error.
> > >>>>
> > >>>> After all, it *is* a poll() bug, so it is only fair that it is tested as
> > >>>> part of the epoll test.
> > >>>>
> > >>>> Can you call EPOLL_CTL_ADD with ev.events set to 0? And then call it again
> > >>>> with the actual value that you need? If that triggers this issue as well,
> > >>>> then that is a nice test (but perhaps EPOLL_CTL_ADD won't call poll() if
> > >>>> ev.events is 0, but perhaps EPOLLERR would work instead of 0).
> > >>>
> > >>> Yup, actually the following is enough to make v4l2-compliance -s fail
> > >>> with vicodec:
> > >>
> > >> Does it also fail with vivid? I am curious to know whether this issue is
> > >> m2m specific or a more general problem.
> > >
> > > It does fail actually! And that made me notice that vb2_poll() uses
> > > the same pattern as v4l2_m2m_poll() (probably because the latter is
> > > inspired by the former?) and needs to be fixed similarly. I will send
> > > another patch to fix vb2_poll() as well, thanks for pointing it out!
> >
> > I was afraid of that.
> >
> > Testing epoll for control events would be interesting as well. The
> > vivid radio device is an example of a device that has controls, but
> > does not do streaming (so is not using vb2).
> >
> > But from what I can see v4l2_ctrl_poll() does the right thing, so this
> > should be fine.
>
> Indeed, it unconditionally calls poll_wait() with all the wait queues
> that may wake us up (that is, only one), so there is no problem there.

Sorry, I noticed that this patch was marked with "Changes Requested"
in patchwork, but isn't it valid as-is? We need a similar change to
VB2, but that should go as a separate patch IMHO. I'm fine with doing
both in one go if you prefer that though.