[PATCH 0/2] media: rockchip: rkcif: ENUM_FMT and bytesperline fixes

[PATCH 0/2] media: rockchip: rkcif: ENUM_FMT and bytesperline fixes

[Faizel K B]

This series fixes two issues in the rkcif MIPI capture path, found while
using the libcamera simple pipeline handler with a Sony IMX415 on a
Radxa ROCK 5B+ (RK3588).

Patch 1 makes VIDIOC_ENUM_FMT honour the mbus_code field, as required
for V4L2_CAP_IO_MC devices. Without it, userspace can select a pixel
format that doesn't match the media bus format, e.g. a different Bayer
order, and the captured images have wrong colours.

Patch 2 aligns bytesperline to 8 bytes, as required by the MIPI DMA,
and honours a larger value requested by userspace. This allows the
capture buffers to be imported as DMA-BUFs by the libcamera GPU
software ISP on Panfrost, which needs a 64-byte stride alignment,
instead of copying every frame.

Tested on ROCK 5B+ with IMX415 (SGBRG10, 3864x2192): correct colours
without any userspace workaround, and a stride of 7936 bytes accepted
and used for zero-copy import. The corresponding libcamera change is
under review at:
https://patchwork.libcamera.org/patch/28662/

Faizel K B (2):
  media: rockchip: rkcif: Filter enumerated formats by media bus code
  media: rockchip: rkcif: Honour the requested bytesperline on MIPI

 .../platform/rockchip/rkcif/rkcif-stream.c    | 37 ++++++++++++++++---
 1 file changed, 32 insertions(+), 5 deletions(-)


base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
-- 
2.43.0

[PATCH 1/2] media: rockchip: rkcif: Filter enumerated formats by media bus code

[Faizel K B]

VIDIOC_ENUM_FMT ignores the mbus_code field and returns all formats,
although the video devices set V4L2_CAP_IO_MC. Return only the formats
that match the requested media bus code.

Otherwise userspace can pick a pixel format that doesn't match the
media bus format, and then interprets the captured data in the wrong
format.

Signed-off-by: Faizel K B <faizel.kb@gmail.com>
---
 .../media/platform/rockchip/rkcif/rkcif-stream.c  | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)

diff --git a/drivers/media/platform/rockchip/rkcif/rkcif-stream.c b/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
index 3130d420ad55..f2a24832a8cf 100644
--- a/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
+++ b/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
@@ -395,13 +395,20 @@ static int rkcif_stream_enum_formats(struct file *file, void *priv,
 				     struct v4l2_fmtdesc *f)
 {
 	struct rkcif_stream *stream = video_drvdata(file);
+	unsigned int i, n = 0;
 
-	if (f->index >= stream->out_fmts_num)
-		return -EINVAL;
+	for (i = 0; i < stream->out_fmts_num; i++) {
+		if (f->mbus_code &&
+		    stream->out_fmts[i].mbus_code != f->mbus_code)
+			continue;
 
-	f->pixelformat = stream->out_fmts[f->index].fourcc;
+		if (n++ == f->index) {
+			f->pixelformat = stream->out_fmts[i].fourcc;
+			return 0;
+		}
+	}
 
-	return 0;
+	return -EINVAL;
 }
 
 static int rkcif_stream_enum_framesizes(struct file *file, void *fh,
-- 
2.43.0

Re: [PATCH 1/2] media: rockchip: rkcif: Filter enumerated formats by media bus code

[Mehdi Djait]

Hello,

On Thu, Nov 05, 2026 at 04:23:09PM -0700, Faizel K B wrote:
> VIDIOC_ENUM_FMT ignores the mbus_code field and returns all formats,
> although the video devices set V4L2_CAP_IO_MC. Return only the formats
> that match the requested media bus code.
> 
> Otherwise userspace can pick a pixel format that doesn't match the
> media bus format, and then interprets the captured data in the wrong
> format.
> 
> Signed-off-by: Faizel K B <faizel.kb@gmail.com>
> ---
>  .../media/platform/rockchip/rkcif/rkcif-stream.c  | 15 +++++++++++----
>  1 file changed, 11 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/media/platform/rockchip/rkcif/rkcif-stream.c b/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
> index 3130d420ad55..f2a24832a8cf 100644
> --- a/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
> +++ b/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
> @@ -395,13 +395,20 @@ static int rkcif_stream_enum_formats(struct file *file, void *priv,
>  				     struct v4l2_fmtdesc *f)
>  {
>  	struct rkcif_stream *stream = video_drvdata(file);
> +	unsigned int i, n = 0;
>  
> -	if (f->index >= stream->out_fmts_num)
> -		return -EINVAL;
> +	for (i = 0; i < stream->out_fmts_num; i++) {
> +		if (f->mbus_code &&
> +		    stream->out_fmts[i].mbus_code != f->mbus_code)

what about dvp capture. The dvp_out_fmts[] does not have the mbus_code
set.

> +			continue;

--
Kind Regards
Mehdi Djait

Re: [PATCH 1/2] media: rockchip: rkcif: Filter enumerated formats by media bus code

[Faizel K B]

Hi Mehdi,

On Mon, Nov 09, 2026 at 01:27:34PM +0100, Mehdi Djait wrote:

>> --- a/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
>> +++ b/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
>> @@ -395,13 +395,20 @@ static int rkcif_stream_enum_formats(struct file *file, void *priv,
>>  				     struct v4l2_fmtdesc *f)
>>  {
>>  	struct rkcif_stream *stream = video_drvdata(file);
>> +	unsigned int i, n = 0;
>>
>> -	if (f->index >= stream->out_fmts_num)
>> -		return -EINVAL;
>> +	for (i = 0; i < stream->out_fmts_num; i++) {
>> +		if (f->mbus_code &&
>> +		    stream->out_fmts[i].mbus_code != f->mbus_code)
>
>what about dvp capture. The dvp_out_fmts[] does not have the mbus_code
>set.
>

Thanks for the review. Yes, the DVP case is missing. How about only   
filtering the formats that have a media bus code, so DVP behaves as
before (see the change).

        const struct rkcif_output_fmt *fmt = &stream->out_fmts[i];

        if (f->mbus_code && fmt->mbus_code &&
            fmt->mbus_code != f->mbus_code)
                continue;

I don't have DVP hardware to test with. I can submit a v2 if that's ok.

Regards,
Faizel

Re: [PATCH 1/2] media: rockchip: rkcif: Filter enumerated formats by media bus code

[Michael Riesch]

Hi there,

On 11/9/26 20:54, Faizel K B wrote:
> Hi Mehdi,
> 
> On Mon, Nov 09, 2026 at 01:27:34PM +0100, Mehdi Djait wrote:
> 
>>> --- a/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
>>> +++ b/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
>>> @@ -395,13 +395,20 @@ static int rkcif_stream_enum_formats(struct
>>> file *file, void *priv,
>>>                       struct v4l2_fmtdesc *f)
>>>  {
>>>      struct rkcif_stream *stream = video_drvdata(file);
>>> +    unsigned int i, n = 0;
>>>
>>> -    if (f->index >= stream->out_fmts_num)
>>> -        return -EINVAL;
>>> +    for (i = 0; i < stream->out_fmts_num; i++) {
>>> +        if (f->mbus_code &&
>>> +            stream->out_fmts[i].mbus_code != f->mbus_code)
>>
>> what about dvp capture. The dvp_out_fmts[] does not have the mbus_code
>> set.
>>
> 
> Thanks for the review. Yes, the DVP case is missing. How about only  
> filtering the formats that have a media bus code, so DVP behaves as
> before (see the change).

This sounds reasonable.

> 
>        const struct rkcif_output_fmt *fmt = &stream->out_fmts[i];
> 
>        if (f->mbus_code && fmt->mbus_code &&
>            fmt->mbus_code != f->mbus_code)
>                continue;
> 
> I don't have DVP hardware to test with. I can submit a v2 if that's ok.

Looking forward to the v2.

Best regards,
Michael

> 
> Regards,
> Faizel
> 
> 

[PATCH 2/2] media: rockchip: rkcif: Honour the requested bytesperline on MIPI

[Faizel K B]

The MIPI DMA uses bytesperline as the line stride, which must be a
multiple of 8 bytes, but v4l2_fill_pixfmt_mp() doesn't align it.
Any larger bytesperline requested by userspace is also replaced with
the minimum.

Align bytesperline to 8 bytes and honour a larger requested value, up
to the limit of the line width register. This allows DMA-BUF importers
that need a larger stride alignment to use the buffers directly.

Signed-off-by: Faizel K B <faizel.kb@gmail.com>
---
 .../platform/rockchip/rkcif/rkcif-stream.c    | 22 ++++++++++++++++++-
 1 file changed, 21 insertions(+), 1 deletion(-)

diff --git a/drivers/media/platform/rockchip/rkcif/rkcif-stream.c b/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
index f2a24832a8cf..e50f20e40f7a 100644
--- a/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
+++ b/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
@@ -23,6 +23,7 @@
 #define CIF_MIN_HEIGHT	 64
 #define CIF_MAX_WIDTH	 8192
 #define CIF_MAX_HEIGHT	 8192
+#define CIF_MAX_BYTESPERLINE 0xffff8 /* VLW: 20 bits, 8-byte aligned */
 
 static inline struct rkcif_buffer *to_rkcif_buffer(struct vb2_v4l2_buffer *vb)
 {
@@ -334,20 +335,39 @@ static const struct vb2_ops rkcif_stream_vb2_ops = {
 	.stop_streaming = rkcif_stream_stop_streaming,
 };
 
+static void rkcif_stream_set_bytesperline(struct v4l2_pix_format_mplane *pix,
+					  u32 req_bpl)
+{
+	u32 min_bpl = ALIGN(pix->plane_fmt[0].bytesperline, 8);
+	u32 max_bpl = min_t(u32, CIF_MAX_BYTESPERLINE,
+			    ALIGN_DOWN(U32_MAX / pix->height, 8));
+	u32 bpl = min_bpl;
+
+	if (req_bpl > min_bpl && req_bpl <= max_bpl)
+		bpl = ALIGN(req_bpl, 8);
+
+	pix->plane_fmt[0].bytesperline = bpl;
+	pix->plane_fmt[0].sizeimage = bpl * pix->height;
+}
+
 static int rkcif_stream_fill_format(struct rkcif_stream *stream,
 				    struct v4l2_pix_format_mplane *pix)
 {
 	const struct rkcif_output_fmt *fmt;
-	u32 height, width;
+	u32 height, width, req_bpl;
 	int ret;
 
 	fmt = rkcif_stream_find_output_fmt(stream, true, pix->pixelformat);
 	height = clamp_t(u32, pix->height, CIF_MIN_HEIGHT, CIF_MAX_HEIGHT);
 	width = clamp_t(u32, pix->width, CIF_MIN_WIDTH, CIF_MAX_WIDTH);
+	req_bpl = pix->plane_fmt[0].bytesperline;
 	ret = v4l2_fill_pixfmt_mp(pix, fmt->fourcc, width, height);
 	if (ret)
 		return ret;
 
+	if (stream->interface->type == RKCIF_IF_MIPI)
+		rkcif_stream_set_bytesperline(pix, req_bpl);
+
 	pix->field = V4L2_FIELD_NONE;
 
 	return 0;
-- 
2.43.0

Re: [PATCH 2/2] media: rockchip: rkcif: Honour the requested bytesperline on MIPI

[Michael Riesch]

Hi there,

Thanks for the patch!

On 11/6/26 00:23, Faizel K B wrote:
> The MIPI DMA uses bytesperline as the line stride, which must be a
> multiple of 8 bytes, but v4l2_fill_pixfmt_mp() doesn't align it.

Come to think of it, TRM p. 2111 says for e.g. VICAP_MIPI0_VLW_ID0:
"Virtual line width of even frame for ID0 path(must be aligned to double
word)." which I would interpret as "it must be a multiple of four
bytes". What do you think?

> Any larger bytesperline requested by userspace is also replaced with
> the minimum.
> 
> Align bytesperline to 8 bytes and honour a larger requested value, up
> to the limit of the line width register. This allows DMA-BUF importers
> that need a larger stride alignment to use the buffers directly.
> 
> Signed-off-by: Faizel K B <faizel.kb@gmail.com>
> ---
>  .../platform/rockchip/rkcif/rkcif-stream.c    | 22 ++++++++++++++++++-
>  1 file changed, 21 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/media/platform/rockchip/rkcif/rkcif-stream.c b/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
> index f2a24832a8cf..e50f20e40f7a 100644
> --- a/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
> +++ b/drivers/media/platform/rockchip/rkcif/rkcif-stream.c
> @@ -23,6 +23,7 @@
>  #define CIF_MIN_HEIGHT	 64
>  #define CIF_MAX_WIDTH	 8192
>  #define CIF_MAX_HEIGHT	 8192
> +#define CIF_MAX_BYTESPERLINE 0xffff8 /* VLW: 20 bits, 8-byte aligned */
>  
>  static inline struct rkcif_buffer *to_rkcif_buffer(struct vb2_v4l2_buffer *vb)
>  {
> @@ -334,20 +335,39 @@ static const struct vb2_ops rkcif_stream_vb2_ops = {
>  	.stop_streaming = rkcif_stream_stop_streaming,
>  };
>  
> +static void rkcif_stream_set_bytesperline(struct v4l2_pix_format_mplane *pix,
> +					  u32 req_bpl)
> +{
> +	u32 min_bpl = ALIGN(pix->plane_fmt[0].bytesperline, 8);

Please add support for the bytesperline field in all the plane_fmt
entries here...

> +	u32 max_bpl = min_t(u32, CIF_MAX_BYTESPERLINE,
> +			    ALIGN_DOWN(U32_MAX / pix->height, 8));
> +	u32 bpl = min_bpl;
> +
> +	if (req_bpl > min_bpl && req_bpl <= max_bpl)
> +		bpl = ALIGN(req_bpl, 8);
> +
> +	pix->plane_fmt[0].bytesperline = bpl;
> +	pix->plane_fmt[0].sizeimage = bpl * pix->height;
> +}
> +
>  static int rkcif_stream_fill_format(struct rkcif_stream *stream,
>  				    struct v4l2_pix_format_mplane *pix)
>  {
>  	const struct rkcif_output_fmt *fmt;
> -	u32 height, width;
> +	u32 height, width, req_bpl;
>  	int ret;
>  
>  	fmt = rkcif_stream_find_output_fmt(stream, true, pix->pixelformat);
>  	height = clamp_t(u32, pix->height, CIF_MIN_HEIGHT, CIF_MAX_HEIGHT);
>  	width = clamp_t(u32, pix->width, CIF_MIN_WIDTH, CIF_MAX_WIDTH);
> +	req_bpl = pix->plane_fmt[0].bytesperline;

... and here.

>  	ret = v4l2_fill_pixfmt_mp(pix, fmt->fourcc, width, height);
>  	if (ret)
>  		return ret;
>  
> +	if (stream->interface->type == RKCIF_IF_MIPI)
> +		rkcif_stream_set_bytesperline(pix, req_bpl);

Why is this restricted to the MIPI interfaces?

Best regards,
Michael

> +
>  	pix->field = V4L2_FIELD_NONE;
>  
>  	return 0;