[PATCH v3 0/2] media: i2c: vd55g1: Two control fixes

[Fernando Rimoli <fernandorimoli11@gmail.com>]

1/2 marks the auto exposure lock and bias controls inactive rather than
grabbed in manual exposure mode. It is unchanged; Sakari's question on
it is answered on v2.

2/2 replaces v2's hand-written event ops, as Sakari asked: it
initialises the controls before v4l2_subdev_init_finalize(), which
enables control events. The same order fixes a locking bug the driver
has had since it was added: the state lock was set before the control
handler lock existed, so the two were separate mutexes. Moving the
finalise alone deadlocks the first stream start, because
vd55g1_grab_ctrls() then takes the control handler lock a second time,
so 2/2 also switches it to __v4l2_ctrl_grab(). 2/2 does not depend on
1/2.

Because of the locking bug, 2/2 carries Fixes and Cc: stable, unlike v2.
Benjamin, drop them if you would rather not have it backported.

Changes since v2:
 - 2/2: new approach, see above (Sakari)

Testing. I do not have a VD55G1, so this was tested on the VD55G0 of a
Surface Pro 9 (IPU6) through an out-of-tree driver derived from vd55g1.
It is identical in the code 2/2 touches, apart from one more
format-dependent control (pixel rate). A test-only print in s_ctrl()
compared the two locks:

 - before 2/2, VIDIOC_S_CTRL ran s_ctrl() with only the control lock
   held, and setting the active format and stream on ran it with only
   the state lock held;
 - with 2/2 they are one lock, held in all three cases;
 - without the __v4l2_ctrl_grab() change, the first stream start hung in
   vd55g1_grab_ctrls(), blocked on the mutex it already held.

The controls at probe and after a format change match the unpatched
driver, apart from the default of that driver's pixel rate control,
which my port of 2/2 left at its placeholder. The flip and test pattern
controls are grabbed while streaming and released after. R8 644x604, R10
and R10_CSI2P 640x480 streamed 90 of 90 frames, and five start/stop
cycles were clean. v4l2-compliance on the subdev gives the same results
as with v2, in both exposure modes, and control events pass. The kernel
here has no lockdep, hence the print.

v2: https://lore.kernel.org/linux-media/20261101195006.220657-1-fernandorimoli11@gmail.com/
v1: https://lore.kernel.org/linux-media/20261025154031.27083-1-fernandorimoli11@gmail.com/

Fernando Rimoli (2):
  media: i2c: vd55g1: Mark the AE lock and bias controls inactive in
    manual mode
  media: i2c: vd55g1: Use the control handler lock for the subdev state

 drivers/media/i2c/vd55g1.c | 65 ++++++++++++++++++++------------------
 1 file changed, 35 insertions(+), 30 deletions(-)

base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
-- 
2.43.0