[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <33b045d6-deb9-2c09-3f74-5ca13f4d2e46@arm.com>
Date: Fri, 29 May 2020 13:47:16 +0100
From: Steven Price <steven.price@....com>
To: Clément Péron <peron.clem@...il.com>,
Robin Murphy <robin.murphy@....com>
Cc: Rob Herring <robh@...nel.org>,
Tomeu Vizoso <tomeu.vizoso@...labora.com>,
Alyssa Rosenzweig <alyssa.rosenzweig@...labora.com>,
Viresh Kumar <vireshk@...nel.org>, Nishanth Menon <nm@...com>,
Stephen Boyd <sboyd@...nel.org>,
Maxime Ripard <mripard@...nel.org>,
Chen-Yu Tsai <wens@...e.org>,
linux-kernel <linux-kernel@...r.kernel.org>,
dri-devel <dri-devel@...ts.freedesktop.org>
Subject: Re: [PATCH 05/15] drm/panfrost: use spinlock instead of atomic
On 29/05/2020 13:35, Clément Péron wrote:
> Hi Robin,
>
> On Fri, 29 May 2020 at 14:20, Robin Murphy <robin.murphy@....com> wrote:
>>
>> On 2020-05-10 17:55, Clément Péron wrote:
>>> Convert busy_count to a simple int protected by spinlock.
>>
>> A little more reasoning might be nice.
>
> I have follow the modification requested for lima devfreq and clearly
> don't have any argument to switch to spinlock.
>
> The Lima Maintainer asked to change witht the following reason :
> "Better make this count a normal int which is also protected by the spinlock,
> because current implementation can't protect atomic ops for state change
> and busy idle check and we are using spinlock already"
>
>>
>>> Signed-off-by: Clément Péron <peron.clem@...il.com>
>>> ---
>> [...]
>>> diff --git a/drivers/gpu/drm/panfrost/panfrost_devfreq.h b/drivers/gpu/drm/panfrost/panfrost_devfreq.h
>>> index 0697f8d5aa34..e6629900a618 100644
>>> --- a/drivers/gpu/drm/panfrost/panfrost_devfreq.h
>>> +++ b/drivers/gpu/drm/panfrost/panfrost_devfreq.h
>>> @@ -4,6 +4,7 @@
>>> #ifndef __PANFROST_DEVFREQ_H__
>>> #define __PANFROST_DEVFREQ_H__
>>>
>>> +#include <linux/spinlock.h>
>>> #include <linux/ktime.h>
>>>
>>> struct devfreq;
>>> @@ -14,10 +15,17 @@ struct panfrost_device;
>>> struct panfrost_devfreq {
>>> struct devfreq *devfreq;
>>> struct thermal_cooling_device *cooling;
>>> +
>>> ktime_t busy_time;
>>> ktime_t idle_time;
>>> ktime_t time_last_update;
>>> - atomic_t busy_count;
>>> + int busy_count;
>>> + /*
>>> + * Protect busy_time, idle_time, time_last_update and busy_count
>>> + * because these can be updated concurrently, for example by the GP
>>> + * and PP interrupts.
>>> + */
>>
>> Nit: this comment is clearly wrong, since we only have Job, GPU and MMU
>> interrupts here. I guess if there is a race it would be between
>> submission/completion/timeout on different job slots.
>
> It's copy/paste from lima I will update it,
Lima ('Utgard') has separate units for geometry and pixel processing
(GP/PP). For Panfrost ('Midgard'/'Bifrost') we don't have that
separation, however there are multiple job slots. which are implemented
as multiple DRM schedulers. So the same fix is appropriate, but clearly
I missed this comment because it's referring to GP/PP which don't exist
for Midgard/Bifrost.
>>
>> Given that, should this actually be considered a fix for 9e62b885f715
>> ("drm/panfrost: Simplify devfreq utilisation tracking")?
>
> I can't say if it can be considered as a fix, I didn't see any
> improvement on my board before and after this patch.
> I'm still facing some issue and didn't have time to fully investigate it.
Technically this is a fix - there's a small race which could cause the
devfreq information to become corrupted. However it would resolve itself
on the next devfreq interval when panfrost_devfreq_reset() is called. So
the impact is very minor (devfreq gets some bogus figures). The
important variable (busy_count) was already an atomic so won't be affected.
Steve
> Thanks for you review,
>
>
>>
>> Robin.
Powered by blists - more mailing lists