v4l2_m2m_job_finish() is typically called when
DMA operations complete, in interrupt handlers or DMA
completion callbacks. Calling .device_run from v4l2_m2m_job_finish
creates a nasty re-entrancy path into the driver.

Moreover, some implementation of .device_run might need to sleep,
as is the case for drivers supporting the Request API,
where controls are applied via v4l2_ctrl_request_setup,
which takes the ctrl handler mutex.

This commit adds a deferred context that calls v4l2_m2m_try_run,
and gets scheduled by v4l2_m2m_job_finish().

Before this change, device_run would be called from these
paths:

vb2_m2m_request_queue, or
v4l2_m2m_streamon, or
v4l2_m2m_qbuf
  v4l2_m2m_try_schedule
    v4l2_m2m_try_run
      .device_run

v4l2_m2m_job_finish
  v4l2_m2m_try_run
    .device_run

After this change, the latter is now gone and instead:

v4l2_m2m_device_run_work
  v4l2_m2m_try_run
    .device_run

Signed-off-by: Ezequiel Garcia <ezequiel@collabora.com>
Signed-off-by: Hans Verkuil <hverkuil-cisco@xs4all.nl>
Signed-off-by: Mauro Carvalho Chehab <mchehab+samsung@kernel.org>