Re: [PATCH 3/3] media: hantro: add per-context fdinfo usage stats

[Nicolas Dufresne <nicolas.dufresne@collabora.com>]

[-- Attachment #1: Type: text/plain, Size: 5139 bytes --]

Hi,

Le vendredi 12 juin 2026 à 10:26 -0400, Detlev Casanova a écrit :
> From: Christopher Healy <healych@amazon.com>
> 
> Add per-file-descriptor hardware utilization tracking to the Hantro
> VPU stateless codec driver, exposed via /proc/<pid>/fdinfo/<fd>.
> 
> Record a ktime timestamp when each job is submitted to hardware in
> device_run(), and accumulate the elapsed nanoseconds when the job
> completes in hantro_job_finish(). Report the accumulated time along
> with the current clock frequency through the new V4L2 show_fdinfo
> callback.
> 
> The output uses a media- key prefix with the following keys:
>   media-driver:           driver name
>   media-engine-<eng>:     accumulated busy time in nanoseconds
>   media-maxfreq-<eng>:    maximum engine frequency in Hz
>   media-curfreq-<eng>:    current engine frequency in Hz
> 
> Where <eng> is "decoder" or "encoder" depending on the context.
> 
> This enables userspace monitoring tools to compute per-process decoder
> and encoder utilization. The current and max frequency keys report the
> same value today since the driver lacks devfreq support, but will
> diverge once DVFS is added, allowing userspace to approximate true
> capacity utilization without any fdinfo code changes. A future series
> can add hardware cycle counter support (via media-cycles-<eng>) for
> exact utilization under DVFS, with no changes to the existing uAPI.
> 
> Signed-off-by: Christopher Healy <healych@amazon.com>
> Signed-off-by: Detlev Casanova <detlev.casanova@collabora.com>
> ---
>  drivers/media/platform/verisilicon/hantro.h     |  5 +++++
>  drivers/media/platform/verisilicon/hantro_drv.c | 22 ++++++++++++++++++++++
>  2 files changed, 27 insertions(+)
> 
> diff --git a/drivers/media/platform/verisilicon/hantro.h b/drivers/media/platform/verisilicon/hantro.h
> index 0353de154a1e..10287e5cebac 100644
> --- a/drivers/media/platform/verisilicon/hantro.h
> +++ b/drivers/media/platform/verisilicon/hantro.h
> @@ -16,6 +16,7 @@
>  #include <linux/videodev2.h>
>  #include <linux/wait.h>
>  #include <linux/clk.h>
> +#include <linux/ktime.h>
>  #include <linux/reset.h>
>  
>  #include <media/v4l2-ctrls.h>
> @@ -268,6 +269,10 @@ struct hantro_ctx {
>  	struct hantro_postproc_ctx postproc;
>  	bool need_postproc;
>  
> +	/* Statistics for debugging and performance measurements. */
> +	ktime_t start_time;
> +	u64 total_ns;
> +
>  	/* Specific for particular codec modes. */
>  	union {
>  		struct hantro_h264_dec_hw_ctx h264_dec;
> diff --git a/drivers/media/platform/verisilicon/hantro_drv.c b/drivers/media/platform/verisilicon/hantro_drv.c
> index 2e81877f640f..139c2e6a83aa 100644
> --- a/drivers/media/platform/verisilicon/hantro_drv.c
> +++ b/drivers/media/platform/verisilicon/hantro_drv.c
> @@ -90,6 +90,8 @@ static void hantro_job_finish(struct hantro_dev *vpu,
>  			      struct hantro_ctx *ctx,
>  			      enum vb2_buffer_state result)
>  {
> +	ctx->total_ns += ktime_to_ns(ktime_sub(ktime_get(), ctx->start_time));
> +
>  	pm_runtime_put_autosuspend(vpu->dev);
>  
>  	clk_bulk_disable(vpu->variant->num_clocks, vpu->clocks);
> @@ -186,6 +188,8 @@ static void device_run(void *priv)
>  
>  	v4l2_m2m_buf_copy_metadata(src, dst);
>  
> +	ctx->start_time = ktime_get();
> +
>  	if (ctx->codec_ops->run(ctx))
>  		goto err_cancel_job;
>  
> @@ -701,10 +705,28 @@ static int hantro_release(struct file *filp)
>  	return 0;
>  }
>  
> +static void hantro_show_fdinfo(struct seq_file *m, struct file *f)
> +{
> +	struct hantro_ctx *ctx = file_to_ctx(f);
> +	struct hantro_dev *vpu = ctx->dev;
> +
> +	seq_printf(m, "media-driver:\t%s\n", DRIVER_NAME);
> +	seq_printf(m, "media-engine-%s:\t%llu ns\n",
> +		   ctx->is_encoder ? "encoder" : "decoder",
> +		   ctx->total_ns);
> +	seq_printf(m, "media-maxfreq-%s:\t%lu Hz\n",
> +		   ctx->is_encoder ? "encoder" : "decoder",
> +		   clk_get_rate(vpu->clocks[0].clk));
> +	seq_printf(m, "media-curfreq-%s:\t%lu Hz\n",
> +		   ctx->is_encoder ? "encoder" : "decoder",
> +		   clk_get_rate(vpu->clocks[0].clk));

Similar to what was done in DRM helpers, I would create a common structure,
maybe in a v4l2-stats library (or some better name). The driver would fill that
structure directly, or through helpers (such as software timer helper, cycle
count translation helper), and finally, this show_fd_info() would be a one
liners, just like DRM do.

This will ensure full consistency in the trace, and avoid possible future
unresolvable compatibility issue due to past mistakes. If we ever let a bug
through (of course we won't :-D) all driver will have the same bug, and it will
be easier to not break userspace.

Nicolas

> +}
> +
>  static const struct v4l2_file_operations hantro_fops = {
>  	.owner = THIS_MODULE,
>  	.open = hantro_open,
>  	.release = hantro_release,
> +	.show_fdinfo = hantro_show_fdinfo,
>  	.poll = v4l2_m2m_fop_poll,
>  	.unlocked_ioctl = video_ioctl2,
>  	.mmap = v4l2_m2m_fop_mmap,

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 228 bytes --]