Re: [PATCH v5 10/16] media: platform: Add mali-c55 3a stats devnode

[Dan Scally <dan.scally@xxxxxxxxxxxxxxxx>]

Hi Laurent

On 16/06/2024 22:19, Laurent Pinchart wrote:
> Hi Dan,
>
> Thank you for the patch.
>
> On Wed, May 29, 2024 at 04:28:52PM +0100, Daniel Scally wrote:
> > Add a new code file to govern the 3a statistics capture node.
> >
> > Acked-by: Nayden Kanchev  <nayden.kanchev@xxxxxxx>
> > Co-developed-by: Jacopo Mondi <jacopo.mondi@xxxxxxxxxxxxxxxx>
> > Signed-off-by: Jacopo Mondi <jacopo.mondi@xxxxxxxxxxxxxxxx>
> > Signed-off-by: Daniel Scally <dan.scally@xxxxxxxxxxxxxxxx>
> > ---
> > Changes in v5:
> >
> > 	- New patch
> >
> >   drivers/media/platform/arm/mali-c55/Makefile  |   3 +-
> >   .../platform/arm/mali-c55/mali-c55-common.h   |  28 ++
> >   .../platform/arm/mali-c55/mali-c55-core.c     |  15 +
> >   .../platform/arm/mali-c55/mali-c55-isp.c      |   1 +
> >   .../arm/mali-c55/mali-c55-registers.h         |   3 +
> >   .../platform/arm/mali-c55/mali-c55-stats.c    | 350 ++++++++++++++++++
> >   6 files changed, 399 insertions(+), 1 deletion(-)
> >   create mode 100644 drivers/media/platform/arm/mali-c55/mali-c55-stats.c
> >
> > diff --git a/drivers/media/platform/arm/mali-c55/Makefile b/drivers/media/platform/arm/mali-c55/Makefile
> > index 77dcb2fbf0f4..cd5a64bf0c62 100644
> > --- a/drivers/media/platform/arm/mali-c55/Makefile
> > +++ b/drivers/media/platform/arm/mali-c55/Makefile
> > @@ -4,6 +4,7 @@ mali-c55-y := mali-c55-capture.o \
> >   	      mali-c55-core.o \
> >   	      mali-c55-isp.o \
> >   	      mali-c55-tpg.o \
> > -	      mali-c55-resizer.o
> > +	      mali-c55-resizer.o \
> > +	      mali-c55-stats.o
> >   
> >   obj-$(CONFIG_VIDEO_MALI_C55) += mali-c55.o
> > diff --git a/drivers/media/platform/arm/mali-c55/mali-c55-common.h b/drivers/media/platform/arm/mali-c55/mali-c55-common.h
> > index 2d0c4d152beb..44119e04009b 100644
> > --- a/drivers/media/platform/arm/mali-c55/mali-c55-common.h
> > +++ b/drivers/media/platform/arm/mali-c55/mali-c55-common.h
> > @@ -79,6 +79,7 @@ enum mali_c55_isp_pads {
> >   	MALI_C55_ISP_PAD_SINK_VIDEO,
> >   	MALI_C55_ISP_PAD_SOURCE,
> >   	MALI_C55_ISP_PAD_SOURCE_BYPASS,
> > +	MALI_C55_ISP_PAD_SOURCE_3A,
> Functions and structures are named with a "stats" suffix, let's call
> this MALI_C55_ISP_PAD_SOURCE_STATS.
>
> >   	MALI_C55_ISP_NUM_PADS,
> >   };
> >   
> > @@ -194,6 +195,28 @@ struct mali_c55_cap_dev {
> >   	bool streaming;
> >   };
> >   
> > +struct mali_c55_stats_buf {
> > +	struct vb2_v4l2_buffer vb;
> > +	spinlock_t lock;
> All locks require a comment to document what they protect. Same below.
>
> > +	unsigned int segments_remaining;
> > +	struct list_head queue;
> > +	bool failed;
> > +};
> > +
> > +struct mali_c55_stats {
> > +	struct mali_c55 *mali_c55;
> > +	struct video_device vdev;
> > +	struct dma_chan *channel;
> > +	struct vb2_queue queue;
> > +	struct media_pad pad;
> > +	struct mutex lock;
> > +
> > +	struct {
> > +		spinlock_t lock;
> > +		struct list_head queue;
> > +	} buffers;
> > +};
> > +
> >   enum mali_c55_config_spaces {
> >   	MALI_C55_CONFIG_PING,
> >   	MALI_C55_CONFIG_PONG,
> > @@ -224,6 +247,7 @@ struct mali_c55 {
> >   	struct mali_c55_isp isp;
> >   	struct mali_c55_resizer resizers[MALI_C55_NUM_RZRS];
> >   	struct mali_c55_cap_dev cap_devs[MALI_C55_NUM_CAP_DEVS];
> > +	struct mali_c55_stats stats;
> >   
> >   	struct list_head contexts;
> >   	enum mali_c55_config_spaces next_config;
> > @@ -245,6 +269,8 @@ int mali_c55_register_resizers(struct mali_c55 *mali_c55);
> >   void mali_c55_unregister_resizers(struct mali_c55 *mali_c55);
> >   int mali_c55_register_capture_devs(struct mali_c55 *mali_c55);
> >   void mali_c55_unregister_capture_devs(struct mali_c55 *mali_c55);
> > +int mali_c55_register_stats(struct mali_c55 *mali_c55);
> > +void mali_c55_unregister_stats(struct mali_c55 *mali_c55);
> >   struct mali_c55_ctx *mali_c55_get_active_context(struct mali_c55 *mali_c55);
> >   void mali_c55_set_plane_done(struct mali_c55_cap_dev *cap_dev,
> >   			     enum mali_c55_planes plane);
> > @@ -262,5 +288,7 @@ mali_c55_isp_fmt_next(const struct mali_c55_isp_fmt *fmt);
> >   bool mali_c55_isp_is_format_supported(unsigned int mbus_code);
> >   #define for_each_mali_isp_fmt(fmt)\
> >   	for ((fmt) = NULL; ((fmt) = mali_c55_isp_fmt_next((fmt)));)
> > +void mali_c55_stats_fill_buffer(struct mali_c55 *mali_c55,
> > +				enum mali_c55_config_spaces cfg_space);
> >   
> >   #endif /* _MALI_C55_COMMON_H */
> > diff --git a/drivers/media/platform/arm/mali-c55/mali-c55-core.c b/drivers/media/platform/arm/mali-c55/mali-c55-core.c
> > index 50caf5ee7474..9ea70010876c 100644
> > --- a/drivers/media/platform/arm/mali-c55/mali-c55-core.c
> > +++ b/drivers/media/platform/arm/mali-c55/mali-c55-core.c
> > @@ -337,6 +337,16 @@ static int mali_c55_create_links(struct mali_c55 *mali_c55)
> >   		}
> >   	}
> >   
> > +	ret = media_create_pad_link(&mali_c55->isp.sd.entity,
> > +			MALI_C55_ISP_PAD_SOURCE_3A,
> > +			&mali_c55->stats.vdev.entity, 0,
> > +			MEDIA_LNK_FL_ENABLED | MEDIA_LNK_FL_IMMUTABLE);
> > +	if (ret) {
> > +		dev_err(mali_c55->dev,
> > +			"failed to link ISP and 3a stats node\n");
> s/3a stats/stats/
>
> > +		goto err_remove_links;
> > +	}
> > +
> >   	return 0;
> >   
> >   err_remove_links:
> > @@ -350,6 +360,7 @@ static void mali_c55_unregister_entities(struct mali_c55 *mali_c55)
> >   	mali_c55_unregister_isp(mali_c55);
> >   	mali_c55_unregister_resizers(mali_c55);
> >   	mali_c55_unregister_capture_devs(mali_c55);
> > +	mali_c55_unregister_stats(mali_c55);
> >   }
> >   
> >   static int mali_c55_register_entities(struct mali_c55 *mali_c55)
> > @@ -372,6 +383,10 @@ static int mali_c55_register_entities(struct mali_c55 *mali_c55)
> >   	if (ret)
> >   		goto err_unregister_entities;
> >   
> > +	ret = mali_c55_register_stats(mali_c55);
> > +	if (ret)
> > +		goto err_unregister_entities;
> > +
> >   	ret = mali_c55_create_links(mali_c55);
> >   	if (ret)
> >   		goto err_unregister_entities;
> > diff --git a/drivers/media/platform/arm/mali-c55/mali-c55-isp.c b/drivers/media/platform/arm/mali-c55/mali-c55-isp.c
> > index ea8b7b866e7a..94876fba3353 100644
> > --- a/drivers/media/platform/arm/mali-c55/mali-c55-isp.c
> > +++ b/drivers/media/platform/arm/mali-c55/mali-c55-isp.c
> > @@ -564,6 +564,7 @@ int mali_c55_register_isp(struct mali_c55 *mali_c55)
> >   	isp->pads[MALI_C55_ISP_PAD_SINK_VIDEO].flags = MEDIA_PAD_FL_SINK;
> >   	isp->pads[MALI_C55_ISP_PAD_SOURCE].flags = MEDIA_PAD_FL_SOURCE;
> >   	isp->pads[MALI_C55_ISP_PAD_SOURCE_BYPASS].flags = MEDIA_PAD_FL_SOURCE;
> > +	isp->pads[MALI_C55_ISP_PAD_SOURCE_3A].flags = MEDIA_PAD_FL_SOURCE;
> >   
> >   	ret = media_entity_pads_init(&sd->entity, MALI_C55_ISP_NUM_PADS,
> >   				     isp->pads);
> > diff --git a/drivers/media/platform/arm/mali-c55/mali-c55-registers.h b/drivers/media/platform/arm/mali-c55/mali-c55-registers.h
> > index cb27abde2aa5..eb3719245ec3 100644
> > --- a/drivers/media/platform/arm/mali-c55/mali-c55-registers.h
> > +++ b/drivers/media/platform/arm/mali-c55/mali-c55-registers.h
> > @@ -68,6 +68,9 @@
> >   #define MALI_C55_VC_START(v)				((v) & 0xffff)
> >   #define MALI_C55_VC_SIZE(v)				(((v) & 0xffff) << 16)
> >   
> > +#define MALI_C55_REG_1024BIN_HIST			0x054a8
> > +#define MALI_C55_1024BIN_HIST_SIZE			4096
> > +
> >   /* Ping/Pong Configuration Space */
> >   #define MALI_C55_REG_BASE_ADDR				0x18e88
> >   #define MALI_C55_REG_BYPASS_0				0x18eac
> > diff --git a/drivers/media/platform/arm/mali-c55/mali-c55-stats.c b/drivers/media/platform/arm/mali-c55/mali-c55-stats.c
> > new file mode 100644
> > index 000000000000..aa40480ed814
> > --- /dev/null
> > +++ b/drivers/media/platform/arm/mali-c55/mali-c55-stats.c
> > @@ -0,0 +1,350 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +/*
> > + * ARM Mali-C55 ISP Driver - 3A Statistics capture device
> > + *
> > + * Copyright (C) 2023 Ideas on Board Oy
> > + */
> > +
> > +#include <linux/dmaengine.h>
> > +#include <linux/media/arm/mali-c55-config.h>
> > +#include <linux/spinlock.h>
> You're missing some headers here, for
>
> container_of()
> dev_err()
> list_*()
> mutex_init()
> strscpy()
> strscpy()
>
> > +
> > +#include <media/media-entity.h>
> > +#include <media/v4l2-dev.h>
> > +#include <media/v4l2-event.h>
> > +#include <media/v4l2-fh.h>
> > +#include <media/v4l2-ioctl.h>
> > +#include <media/videobuf2-core.h>
> > +#include <media/videobuf2-dma-contig.h>
> > +
> > +#include "mali-c55-common.h"
> > +#include "mali-c55-registers.h"
> > +
> > +static unsigned int metering_space_addrs[] = {
> const
>
> > +	[MALI_C55_CONFIG_PING] = 0x095AC,
> > +	[MALI_C55_CONFIG_PONG] = 0x2156C,
> Lower-case hex constants.
>
> > +};
> > +
> > +static int mali_c55_stats_enum_fmt_meta_cap(struct file *file, void *fh,
> > +					    struct v4l2_fmtdesc *f)
> > +{
> > +	if (f->index || f->type != V4L2_BUF_TYPE_META_CAPTURE)
> > +		return -EINVAL;
> > +
> > +	f->pixelformat = V4L2_META_FMT_MALI_C55_3A_STATS;
> The format could be called V4L2_META_FMT_MALI_C55_STATS. While most
> statistics are related to one of the 3A algorithms, I think it would be
> better to name this generically. It's name bikeshedding only of course.
>
> > +
> > +	return 0;
> > +}
> > +
> > +static int mali_c55_stats_g_fmt_meta_cap(struct file *file, void *fh,
> > +					 struct v4l2_format *f)
> > +{
> > +	static const struct v4l2_meta_format mfmt = {
> > +		.dataformat = V4L2_META_FMT_MALI_C55_3A_STATS,
> > +		.buffersize = sizeof(struct mali_c55_stats_buffer)
> > +	};
> > +
> > +	if (f->type != V4L2_BUF_TYPE_META_CAPTURE)
> > +		return -EINVAL;
> > +
> > +	f->fmt.meta = mfmt;
> > +
> > +	return 0;
> > +}
> > +
> > +static int mali_c55_stats_querycap(struct file *file,
> > +				   void *priv, struct v4l2_capability *cap)
> > +{
> > +	strscpy(cap->driver, MALI_C55_DRIVER_NAME, sizeof(cap->driver));
> > +	strscpy(cap->card, "ARM Mali-C55 ISP", sizeof(cap->card));
> > +
> > +	return 0;
> > +}
> > +
> > +static const struct v4l2_ioctl_ops mali_c55_stats_v4l2_ioctl_ops = {
> > +	.vidioc_reqbufs = vb2_ioctl_reqbufs,
> > +	.vidioc_querybuf = vb2_ioctl_querybuf,
> > +	.vidioc_create_bufs = vb2_ioctl_create_bufs,
> > +	.vidioc_qbuf = vb2_ioctl_qbuf,
> > +	.vidioc_expbuf = vb2_ioctl_expbuf,
> > +	.vidioc_dqbuf = vb2_ioctl_dqbuf,
> > +	.vidioc_prepare_buf = vb2_ioctl_prepare_buf,
> > +	.vidioc_streamon = vb2_ioctl_streamon,
> > +	.vidioc_streamoff = vb2_ioctl_streamoff,
> > +	.vidioc_enum_fmt_meta_cap = mali_c55_stats_enum_fmt_meta_cap,
> > +	.vidioc_g_fmt_meta_cap = mali_c55_stats_g_fmt_meta_cap,
> > +	.vidioc_s_fmt_meta_cap = mali_c55_stats_g_fmt_meta_cap,
> > +	.vidioc_try_fmt_meta_cap = mali_c55_stats_g_fmt_meta_cap,
> > +	.vidioc_querycap = mali_c55_stats_querycap,
> > +	.vidioc_subscribe_event = v4l2_ctrl_subscribe_event,
> > +	.vidioc_unsubscribe_event = v4l2_event_unsubscribe,
> > +};
> > +
> > +static const struct v4l2_file_operations mali_c55_stats_v4l2_fops = {
> > +	.owner = THIS_MODULE,
> > +	.unlocked_ioctl = video_ioctl2,
> > +	.open = v4l2_fh_open,
> > +	.release = vb2_fop_release,
> > +	.poll = vb2_fop_poll,
> > +	.mmap = vb2_fop_mmap,
> > +};
> > +
> > +static int
> > +mali_c55_stats_queue_setup(struct vb2_queue *q, unsigned int *num_buffers,
> > +			   unsigned int *num_planes, unsigned int sizes[],
> > +			   struct device *alloc_devs[])
> > +{
> > +	struct mali_c55_stats *stats = vb2_get_drv_priv(q);
> > +
> > +	if (*num_planes && *num_planes > 1)
> > +		return -EINVAL;
> > +
> > +	if (sizes[0] && sizes[0] != sizeof(struct mali_c55_stats_buffer))
> > +		return -EINVAL;
> > +
> > +	*num_planes = 1;
> > +	sizes[0] = sizeof(struct mali_c55_stats_buffer);
> > +
> > +	if (stats->channel)
> > +		alloc_devs[0] = stats->channel->device->dev;
> > +
> > +	return 0;
> > +}
> > +
> > +static void mali_c55_stats_buf_queue(struct vb2_buffer *vb)
> > +{
> > +	struct mali_c55_stats *stats = vb2_get_drv_priv(vb->vb2_queue);
> > +	struct vb2_v4l2_buffer *vbuf = to_vb2_v4l2_buffer(vb);
> > +	struct mali_c55_stats_buf *buf = container_of(vbuf,
> > +						struct mali_c55_stats_buf, vb);
> > +
> > +	vb2_set_plane_payload(vb, 0, sizeof(struct mali_c55_stats_buffer));
> > +	buf->segments_remaining = 2;
> > +	buf->failed = false;
> > +
> > +	spin_lock(&stats->buffers.lock);
> Isn't the DMA completion handler run from IRQ context ? If so you'll need to use spin_lock_irq() here and in the other function that are not
> called with interrupts disabled.


They're run in the bottom half of the interrupt handler; I'm under the impression that that means 
the interrupts aren't disabled, and it's safe to do...is that mistaken?

> > +	list_add_tail(&buf->queue, &stats->buffers.queue);
> > +	spin_unlock(&stats->buffers.lock);
> > +}
> > +
> > +static void mali_c55_stats_stop_streaming(struct vb2_queue *q)
> > +{
> > +	struct mali_c55_stats *stats = vb2_get_drv_priv(q);
> > +	struct mali_c55_stats_buf *buf, *tmp;
> > +
> > +	spin_lock(&stats->buffers.lock);
> > +
> > +	list_for_each_entry_safe(buf, tmp, &stats->buffers.queue, queue) {
> > +		list_del(&buf->queue);
> > +		vb2_buffer_done(&buf->vb.vb2_buf, VB2_BUF_STATE_ERROR);
> > +	}
> > +
> > +	spin_unlock(&stats->buffers.lock);
> > +}
> > +
> > +static const struct vb2_ops mali_c55_stats_vb2_ops = {
> > +	.queue_setup = mali_c55_stats_queue_setup,
> > +	.buf_queue = mali_c55_stats_buf_queue,
> > +	.wait_prepare = vb2_ops_wait_prepare,
> > +	.wait_finish = vb2_ops_wait_finish,
> > +	.stop_streaming = mali_c55_stats_stop_streaming,
> > +};
> > +
> > +static void
> > +mali_c55_stats_metering_complete(void *param,
> > +				 const struct dmaengine_result *result)
> > +{
> > +	struct mali_c55_stats_buf *buf = param;
> > +
> > +	spin_lock(&buf->lock);
> I wonder if this is needed. Can the DMA engine call the completion
> handlers of two sequential DMA transfers in parallel ?


The DMA engine that's on the system we have can't...I wasn't sure whether that was generically true.

> > +
> > +	if (buf->failed)
> > +		goto out_unlock;
> > +
> > +	buf->vb.vb2_buf.timestamp = ktime_get_boottime_ns();
> > +
> > +	if (result->result != DMA_TRANS_NOERROR) {
> > +		buf->failed = true;
> > +		vb2_buffer_done(&buf->vb.vb2_buf, VB2_BUF_STATE_ERROR);
> This will possibly return the buffer to userspace after the first DMA
> transfer. Userspace could then requeue the buffer to the kernel before
> the completion of the second DMA transfer. That will cause trouble. I
> think you should instead do something like
>
> 	spin_lock(&buf->lock);
>
> 	if (result->result != DMA_TRANS_NOERROR)
> 		buf->failed = true;
>
> 	if (!--buf->segments_remaining) {
> 		buf->vb.vb2_buf.timestamp = ktime_get_boottime_ns();
> 		vb2_buffer_done(&buf->vb.vb2_buf, buf->failed ?
> 				VB2_BUF_STATE_ERROR : VB2_BUF_STATE_DONE);
> 	}
>
> 	spin_unlock(&buf->lock);
>
> The
>
> 	buf->vb.vb2_buf.timestamp = ktime_get_boottime_ns();
>
> line could also be moved to mali_c55_stats_fill_buffer(), which would
> make sure the timestamp is filled in case of DMA submission failures.


Okedokey

> > +		goto out_unlock;
> > +	}
> > +
> > +	if (!--buf->segments_remaining)
> > +		vb2_buffer_done(&buf->vb.vb2_buf, VB2_BUF_STATE_DONE);
> > +
> > +out_unlock:
> > +	spin_unlock(&buf->lock);
> > +}
> > +
> > +static int mali_c55_stats_dma_xfer(struct mali_c55_stats *stats, dma_addr_t src,
> > +				   dma_addr_t dst,
> > +				   struct mali_c55_stats_buf *buf,
> > +				   size_t length,
> > +				   void (*callback)(void *, const struct dmaengine_result *result))
> The same callback is used for both invocations of this function, you can
> drop the parameter and hardcode it below.
Yeah, not even sure now why I had a parameter.
> > +{
> > +	struct dma_async_tx_descriptor *tx;
> > +	dma_cookie_t cookie;
> > +
> > +	tx = dmaengine_prep_dma_memcpy(stats->channel, dst, src, length, 0);
> > +	if (!tx) {
> > +		dev_err(stats->mali_c55->dev, "failed to prep stats DMA\n");
> > +		return -EIO;
> > +	}
> > +
> > +	tx->callback_result = callback;
> > +	tx->callback_param = buf;
> > +
> > +	cookie = dmaengine_submit(tx);
> > +	if (dma_submit_error(cookie)) {
> > +		dev_err(stats->mali_c55->dev, "failed to submit stats DMA\n");
> > +		return -EIO;
> > +	}
> > +
> > +	dma_async_issue_pending(stats->channel);
> > +	return 0;
> > +}
> > +
> > +void mali_c55_stats_fill_buffer(struct mali_c55 *mali_c55,
> > +				enum mali_c55_config_spaces cfg_space)
> > +{
> > +	struct mali_c55_ctx *ctx = mali_c55_get_active_context(mali_c55);
> > +	struct mali_c55_stats *stats = &mali_c55->stats;
> > +	struct mali_c55_stats_buf *buf = NULL;
> > +	dma_addr_t src, dst;
> > +	int ret;
> > +
> > +	spin_lock(&stats->buffers.lock);
> > +	if (!list_empty(&stats->buffers.queue)) {
> > +		buf = list_first_entry(&stats->buffers.queue,
> > +				       struct mali_c55_stats_buf, queue);
> > +		list_del(&buf->queue);
> > +	}
> > +	spin_unlock(&stats->buffers.lock);
> > +
> > +	if (!buf)
> > +		return;
> > +
> > +	buf->vb.sequence = mali_c55->isp.frame_sequence;
> > +
> > +	/*
> > +	 * There are infact two noncontiguous sections of the ISP's
> s/infact/in fact/
>
> > +	 * memory space that hold statistics for 3a algorithms to use. A
> s/use. A/use: a/
>
> > +	 * section in each config space and a global section holding
> > +	 * histograms which is double buffered and so holds data for the
> > +	 * last frame. We need to read both.
> > +	 */
> > +	src = ctx->base + MALI_C55_REG_1024BIN_HIST;
> > +	dst = vb2_dma_contig_plane_dma_addr(&buf->vb.vb2_buf, 0);
> > +
> > +	ret = mali_c55_stats_dma_xfer(stats, src, dst, buf,
> > +				      MALI_C55_1024BIN_HIST_SIZE,
> > +				      mali_c55_stats_metering_complete);
> > +	if (ret)
> > +		goto err_fail_buffer;
> > +
> > +	src = ctx->base + metering_space_addrs[cfg_space];
> > +	dst += MALI_C55_1024BIN_HIST_SIZE;
> > +
> > +	ret = mali_c55_stats_dma_xfer(
> > +		stats, src, dst, buf,
> > +		sizeof(struct mali_c55_stats_buffer) - MALI_C55_1024BIN_HIST_SIZE,
> > +		mali_c55_stats_metering_complete);
> > +	if (ret) {
> > +		dmaengine_terminate_sync(stats->channel);
> > +		goto err_fail_buffer;
> > +	}
> I think you will need to terminate DMA transfers at stream off time.
>
> > +
> > +	return;
> > +
> > +err_fail_buffer:
> > +	vb2_buffer_done(&buf->vb.vb2_buf, VB2_BUF_STATE_ERROR);
> > +}
> > +
> > +void mali_c55_unregister_stats(struct mali_c55 *mali_c55)
> > +{
> > +	struct mali_c55_stats *stats = &mali_c55->stats;
> > +
> > +	if (!video_is_registered(&stats->vdev))
> > +		return;
> > +
> > +	vb2_video_unregister_device(&stats->vdev);
> > +	media_entity_cleanup(&stats->vdev.entity);
> > +	dma_release_channel(stats->channel);
> > +	mutex_destroy(&stats->lock);
> > +}
> > +
> > +int mali_c55_register_stats(struct mali_c55 *mali_c55)
> > +{
> > +	struct mali_c55_stats *stats = &mali_c55->stats;
> > +	struct video_device *vdev = &stats->vdev;
> > +	struct vb2_queue *vb2q = &stats->queue;
> > +	dma_cap_mask_t mask;
> > +	int ret;
> > +
> > +	mutex_init(&stats->lock);
> > +	INIT_LIST_HEAD(&stats->buffers.queue);
> > +
> > +	dma_cap_zero(mask);
> > +	dma_cap_set(DMA_MEMCPY, mask);
> > +
> > +	stats->channel = dma_request_channel(mask, 0, NULL);
> Do we need a CPU fallback in case no DMA is available ?


Yes, actually.

> I'm still very curious to know how long it takes to perform the DMA
> transfer, compared to copying the data with the CPU, and especially
> compared to the frame duration.


On my list of things to test and report :)

> > +	if (!stats->channel) {
> > +		ret = -ENODEV;
> > +		goto err_destroy_mutex;
> > +	}
> > +
> > +	stats->pad.flags = MEDIA_PAD_FL_SINK;
> > +	ret = media_entity_pads_init(&stats->vdev.entity, 1, &stats->pad);
> > +	if (ret)
> > +		goto err_release_dma_channel;
> > +
> > +	vb2q->type = V4L2_BUF_TYPE_META_CAPTURE;
> > +	vb2q->io_modes = VB2_MMAP | VB2_DMABUF;
> > +	vb2q->drv_priv = stats;
> > +	vb2q->mem_ops = &vb2_dma_contig_memops;
> > +	vb2q->ops = &mali_c55_stats_vb2_ops;
> > +	vb2q->buf_struct_size = sizeof(struct mali_c55_stats_buf);
> > +	vb2q->min_queued_buffers = 1;
> > +	vb2q->timestamp_flags = V4L2_BUF_FLAG_TIMESTAMP_MONOTONIC;
> > +	vb2q->lock = &stats->lock;
> > +	vb2q->dev = mali_c55->dev;
> That's not the right device. The device that performs the DMA operation
> is the DMA engine, and that's what you need to pass to vb2. Otherwise
> the DMA address returned by vb2_dma_contig_plane_dma_addr() will be
> mapped to the ISP device, not the DMA engine. In practice, if neither
> are behind an IOMMU, things will likely work, but when that's not the
> case, run into problems.
>
> > +
> > +	ret = vb2_queue_init(vb2q);
> > +	if (ret) {
> > +		dev_err(mali_c55->dev, "stats vb2 queue init failed\n");
> > +		goto err_cleanup_entity;
> > +	}
> > +
> > +	strscpy(stats->vdev.name, "mali-c55 3a stats", sizeof(stats->vdev.name));
> s/3a //
>
> > +	vdev->release = video_device_release_empty;
> That's never right. You should refcount the data structures to ensure
> proper lifetime management.
>
> > +	vdev->fops = &mali_c55_stats_v4l2_fops;
> > +	vdev->ioctl_ops = &mali_c55_stats_v4l2_ioctl_ops;
> > +	vdev->lock = &stats->lock;
> > +	vdev->v4l2_dev = &mali_c55->v4l2_dev;
> > +	vdev->queue = &stats->queue;
> > +	vdev->device_caps = V4L2_CAP_META_CAPTURE | V4L2_CAP_STREAMING;
> > +	vdev->vfl_dir = VFL_DIR_RX;
> > +	video_set_drvdata(vdev, stats);
> > +
> > +	ret = video_register_device(vdev, VFL_TYPE_VIDEO, -1);
> > +	if (ret) {
> > +		dev_err(mali_c55->dev,
> > +			"failed to register stats video device\n");
> > +		goto err_release_vb2q;
> > +	}
> > +
> > +	stats->mali_c55 = mali_c55;
> > +
> > +	return 0;
> > +
> > +err_release_vb2q:
> > +	vb2_queue_release(vb2q);
> > +err_cleanup_entity:
> > +	media_entity_cleanup(&stats->vdev.entity);
> > +err_release_dma_channel:
> > +	dma_release_channel(stats->channel);
> > +err_destroy_mutex:
> > +	mutex_destroy(&stats->lock);
> > +
> > +	return ret;
> > +}