[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <5a52797c-eb03-3dcc-932f-14a08828799e@arm.com>
Date: Wed, 13 Jan 2021 15:35:37 +0000
From: Steven Price <steven.price@....com>
To: Lukasz Luba <lukasz.luba@....com>, linux-kernel@...r.kernel.org,
airlied@...ux.ie, daniel@...ll.ch, dri-devel@...ts.freedesktop.org
Cc: robh@...nel.org, tomeu.vizoso@...labora.com,
alyssa.rosenzweig@...labora.com
Subject: Re: [PATCH] drm/panfrost: Use delayed timer as default in devfreq
profile
On 05/01/2021 16:41, Lukasz Luba wrote:
> Devfreq framework supports 2 modes for monitoring devices.
> Use delayed timer as default instead of deferrable timer
> in order to monitor the GPU status regardless of CPU idle.
>
> Signed-off-by: Lukasz Luba <lukasz.luba@....com>
Looks reasonable to me.
Reviewed-by: Steven Price <steven.price@....com>
I'll apply to drm-misc-next.
> ---
> Hi all,
>
> This is just a simple change but has impact on reliable polling mode
> with guaranteed sampling rate. The devfreq framework has 2 modes for
> checking the device status which updates the statistics. When the
> default mode is used (deferred work) it could happen that the
> framework won't check the device, if the CPU which was running devfreq
> code went idle. In deferred mode, the deferred timer is not fired,
> thus framework won't check the device status.
>
> The log below is taken when glxgears was running. It can be seen that
> periodic checks are missing (should be every 50ms).
>
> ----------------------------------------------------------------------
>
> [ 128.046555] panfrost ff9a0000.gpu: busy 18029406 total 67940380 26 % freq 200 MHz
> [ 128.110538] panfrost ff9a0000.gpu: busy 7975930 total 63980997 12 % freq 200 MHz
> [ 128.210606] panfrost ff9a0000.gpu: busy 21148792 total 100074440 21 % freq 200 MHz
> [ 128.594638] panfrost ff9a0000.gpu: busy 53739467 total 384032219 13 % freq 200 MHz
> [ 129.178749] panfrost ff9a0000.gpu: busy 89206245 total 584111500 15 % freq 200 MHz
> [ 129.294601] panfrost ff9a0000.gpu: busy 16446118 total 115852915 14 % freq 200 MHz
> [ 129.762616] panfrost ff9a0000.gpu: busy 71481587 total 468015722 15 % freq 200 MHz
> [ 129.874762] panfrost ff9a0000.gpu: busy 15614112 total 112139938 13 % freq 200 MHz
> [ 130.026585] panfrost ff9a0000.gpu: busy 21397620 total 151821546 14 % freq 200 MHz
> [ 130.110536] panfrost ff9a0000.gpu: busy 21429365 total 83952016 25 % freq 200 MHz
> [ 130.178524] panfrost ff9a0000.gpu: busy 16458197 total 67989072 24 % freq 200 MHz
> [ 130.246528] panfrost ff9a0000.gpu: busy 9876225 total 68004030 14 % freq 200 MHz
> [ 130.310529] panfrost ff9a0000.gpu: busy 9211725 total 64000524 14 % freq 200 MHz
> [ 130.410528] panfrost ff9a0000.gpu: busy 21780828 total 99999250 21 % freq 200 MHz
>
> ----------------------------------------------------------------------
>
> It has impact on devfreq cooling and thermal governor. They rely on these
> statistics, but sometimes the data is stale.
>
> This patch switches to more reliable devfreq mode, which uses delayed work.
>
> Regards,
> Lukasz
>
>
> drivers/gpu/drm/panfrost/panfrost_devfreq.c | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/drivers/gpu/drm/panfrost/panfrost_devfreq.c b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
> index f44d28fad085..56b3f5935703 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_devfreq.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
> @@ -76,6 +76,7 @@ static int panfrost_devfreq_get_dev_status(struct device *dev,
> }
>
> static struct devfreq_dev_profile panfrost_devfreq_profile = {
> + .timer = DEVFREQ_TIMER_DELAYED,
> .polling_ms = 50, /* ~3 frames */
> .target = panfrost_devfreq_target,
> .get_dev_status = panfrost_devfreq_get_dev_status,
>
Powered by blists - more mailing lists