Re: [PATCH v10 02/15] media: ov8858: support 19.2 MHz clock and manual white balance

From: Maurizio Casciano

Date: Sat Nov 21 2026 - 08:32:52 EST


Hi Sakari and Laurent,

Thanks for the review. I split the OV8858 clock support out into a
local test patch.

> Any others notices that should be added?

The 2017 Intel notice in v10 came from the OV5670 manual white-balance
code, which I dropped from the clock-only change. The retained OV8858
register table matches 18 of 23 address/value pairs in Intel's GPLv2
OV8858 initialization table from the Yoga Book vendor kernel [1].
One more pair, 0x3f08 = 0x08, is also present in the existing Linux
OV8858 driver. I can trace the four changed PLL values (0x0300, 0x0302,
0x030b and 0x030d) back to the first public Yoga Book patch. Their
earlier source is still unverified. The revised clock-only patch
cites [1], carries its 2014 Intel OV8858 notice and records where
the four adjusted values first appeared. I removed the notice
that came with the dropped OV5670 manual white-balance code.

> The Assisted-by: tags are no longer used, please remove them.

I removed those trailers from the local series and kept the description
of Codex's contribution in the clock patch commit message.

> The 24 MHz configurations are lane number (and variant) dependent.

The only 19.2 MHz configuration I have is for an OV8858 R2A with four
data lanes and a 360 MHz link frequency. The test patch rejects other
revision and lane combinations at 19.2 MHz. At 24 MHz it retains the
existing register values for R1A with two lanes and R2A with two or four
lanes. I compared the table writes and final values for both modes in
each configuration. The clock writes move later in the sequence, and
24 MHz still needs a hardware test.

The Media CI review of v10 patch 02 flagged the same issue [2]. The
unconditional 19.2 MHz table selection in v10 could replace the PLL
settings chosen for a two-lane sensor. The revision and lane check in
the local patch prevents that override.

Laurent wrote:
> That's indeed not a good UAPI.

I dropped the red and blue balance controls from both sensor changes
in the local series. The existing scalar digital-gain control remains.
Any proposal for per-component gain can be reviewed separately.

On the X91L with the local 7.4.0-rc1 integration kernel, raw capture
produced 30 distinct frames on each pass in a front, rear, front test.
The rear OV8858 ran with four lanes at a 360 MHz link frequency. On
the front capture node, v4l2-compliance reported 40 passes, 7 failures
and no warnings. The same seven checks failed with the same diagnostics
on the 7.2 baseline using the same node, camera and tool version.
I repeated the capture sequence with the AtomISP and WV517S fixes in
place after retrying the front sensor probe.
On the exact revised 7.4 image, OV8858 bound automatically and
captured 30 distinct rear RAW frames before and after lens rebind.

[1] https://github.com/jekhor/yogabook-linux-android-kernel/blob/574bae692716f1b14093497bfab8a007fe8e460b/drivers/external_drivers/camera/drivers/media/i2c/ov8858.h
[2] https://linuxtv.org/mailman3/hyperkitty/list/media-ci@xxxxxxxxxxx/message/PWB2NAGYQ7EZVEDZNXK3R3UKZ7GGYAQY/

With Best Regards,
Maurizio Casciano