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

[Hans Verkuil]

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.

Regards,

	Hans

> 
> diff --git a/utils/v4l2-compliance/v4l2-test-buffers.cpp
> b/utils/v4l2-compliance/v4l2-test-buffers.cpp
> index 8000db23..b63326cd 100644
> --- a/utils/v4l2-compliance/v4l2-test-buffers.cpp
> +++ b/utils/v4l2-compliance/v4l2-test-buffers.cpp
> @@ -903,6 +903,10 @@ static int captureBufs(struct node *node, struct
> node *node_m2m_cap, const cv4l_
>                 epollfd = epoll_create1(0);
> 
>                 fail_on_test(epollfd < 0);
> +
> +               ev.events = 0;
> +               fail_on_test(epoll_ctl(epollfd, EPOLL_CTL_ADD,
> node->g_fd(), &ev));
> +
>                 if (node->is_m2m)
>                         ev.events = EPOLLIN | EPOLLOUT | EPOLLPRI;
>                 else if (v4l_type_is_output(q.g_type()))
> @@ -910,7 +914,7 @@ static int captureBufs(struct node *node, struct
> node *node_m2m_cap, const cv4l_
>                 else
>                         ev.events = EPOLLIN;
>                 ev.data.fd = node->g_fd();
> -               fail_on_test(epoll_ctl(epollfd, EPOLL_CTL_ADD,
> node->g_fd(), &ev));
> +               fail_on_test(epoll_ctl(epollfd, EPOLL_CTL_MOD,
> node->g_fd(), &ev));
>         }
> 
>         if (pollmode)
> 
>>
>> The epoll_wait() will fail when this issue hits, so that's a good place
>> to add comments explaining this problem.
>>
>> There is one other place where this needs to be tested: testEvents() in
>> v4l2-test-controls.cpp: currently this only tests select(), but there
>> should be a second epoll test here as well that just tests EPOLLPRI.
>>
>> This would catch drivers that do not stream (i.e. no EPOLLIN/OUT) but
>> that do have controls (so support EPOLLPRI).
> 
> I'll take a look there as well, and think about a proper comment
> before sending a patch towards you.
> 
> Cheers,
> Alex.
>