These are a bunch of devfreq fixes for panfrost that came up in a
discussion with Robin Murphy during the code-review of the lima
devfreq patches: [0]
I am only able to test patch #1 properly because the only boards with
panfrost GPU that I have are using an Amlogic SoC. We don't have
support for the OPP tables or dynamic clock changes there yet.
So patches #2 and #3 are compile-tested only.
[0] https://patchwork.freedesktop.org/patch/346898/
Martin Blumenstingl (3):
drm/panfrost: enable devfreq based the "operating-points-v2" property
drm/panfrost: call dev_pm_opp_of_remove_table() in all error-paths
drm/panfrost: Use the mali-supply regulator for control again
drivers/gpu/drm/panfrost/panfrost_devfreq.c | 33 ++++++++++++++++++---
drivers/gpu/drm/panfrost/panfrost_device.h | 1 +
2 files changed, 30 insertions(+), 4 deletions(-)
--
2.24.1
If devfreq_recommended_opp() fails we need to undo
dev_pm_opp_of_add_table() by calling dev_pm_opp_of_remove_table() (just
like we do it in the other error-path below).
Fixes: f3ba91228e8e91 ("drm/panfrost: Add initial panfrost driver")
Signed-off-by: Martin Blumenstingl <[email protected]>
---
drivers/gpu/drm/panfrost/panfrost_devfreq.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/panfrost/panfrost_devfreq.c b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
index 1471588763ce..170f6c8c9651 100644
--- a/drivers/gpu/drm/panfrost/panfrost_devfreq.c
+++ b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
@@ -93,8 +93,10 @@ int panfrost_devfreq_init(struct panfrost_device *pfdev)
cur_freq = clk_get_rate(pfdev->clock);
opp = devfreq_recommended_opp(dev, &cur_freq, 0);
- if (IS_ERR(opp))
+ if (IS_ERR(opp)) {
+ dev_pm_opp_of_remove_table(dev);
return PTR_ERR(opp);
+ }
panfrost_devfreq_profile.initial_freq = cur_freq;
dev_pm_opp_put(opp);
--
2.24.1
dev_pm_opp_set_rate() needs a reference to the regulator which should be
updated when updating the GPU frequency. The name of the regulator has
to be passed at initialization-time using dev_pm_opp_set_regulators().
Add the call to dev_pm_opp_set_regulators() so dev_pm_opp_set_rate()
will update the GPU regulator when updating the frequency (just like
we did this manually before when we open-coded dev_pm_opp_set_rate()).
Fixes: 221bc77914cbcc ("drm/panfrost: Use generic code for devfreq")
Reported-by: Robin Murphy <[email protected]>
Signed-off-by: Martin Blumenstingl <[email protected]>
---
drivers/gpu/drm/panfrost/panfrost_devfreq.c | 22 ++++++++++++++++++++-
drivers/gpu/drm/panfrost/panfrost_device.h | 1 +
2 files changed, 22 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/panfrost/panfrost_devfreq.c b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
index 170f6c8c9651..4f7999c7b44c 100644
--- a/drivers/gpu/drm/panfrost/panfrost_devfreq.c
+++ b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
@@ -74,6 +74,7 @@ static struct devfreq_dev_profile panfrost_devfreq_profile = {
int panfrost_devfreq_init(struct panfrost_device *pfdev)
{
int ret;
+ struct opp_table *opp_table;
struct dev_pm_opp *opp;
unsigned long cur_freq;
struct device *dev = &pfdev->pdev->dev;
@@ -84,9 +85,24 @@ int panfrost_devfreq_init(struct panfrost_device *pfdev)
/* Optional, continue without devfreq */
return 0;
+ opp_table = dev_pm_opp_set_regulators(dev,
+ (const char *[]){ "mali" },
+ 1);
+ if (IS_ERR(opp_table)) {
+ ret = PTR_ERR(opp_table);
+
+ /* Continue if the optional regulator is missing */
+ if (ret != -ENODEV)
+ return ret;
+ } else {
+ pfdev->devfreq.regulators_opp_table = opp_table;
+ }
+
ret = dev_pm_opp_of_add_table(dev);
- if (ret)
+ if (ret) {
+ dev_pm_opp_put_regulators(pfdev->devfreq.regulators_opp_table);
return ret;
+ }
panfrost_devfreq_reset(pfdev);
@@ -95,6 +111,7 @@ int panfrost_devfreq_init(struct panfrost_device *pfdev)
opp = devfreq_recommended_opp(dev, &cur_freq, 0);
if (IS_ERR(opp)) {
dev_pm_opp_of_remove_table(dev);
+ dev_pm_opp_put_regulators(pfdev->devfreq.regulators_opp_table);
return PTR_ERR(opp);
}
@@ -106,6 +123,7 @@ int panfrost_devfreq_init(struct panfrost_device *pfdev)
if (IS_ERR(devfreq)) {
DRM_DEV_ERROR(dev, "Couldn't initialize GPU devfreq\n");
dev_pm_opp_of_remove_table(dev);
+ dev_pm_opp_put_regulators(pfdev->devfreq.regulators_opp_table);
return PTR_ERR(devfreq);
}
pfdev->devfreq.devfreq = devfreq;
@@ -124,6 +142,8 @@ void panfrost_devfreq_fini(struct panfrost_device *pfdev)
if (pfdev->devfreq.cooling)
devfreq_cooling_unregister(pfdev->devfreq.cooling);
dev_pm_opp_of_remove_table(&pfdev->pdev->dev);
+ if (pfdev->devfreq.regulators_opp_table)
+ dev_pm_opp_put_regulators(pfdev->devfreq.regulators_opp_table);
}
void panfrost_devfreq_resume(struct panfrost_device *pfdev)
diff --git a/drivers/gpu/drm/panfrost/panfrost_device.h b/drivers/gpu/drm/panfrost/panfrost_device.h
index 06713811b92c..4878b239e301 100644
--- a/drivers/gpu/drm/panfrost/panfrost_device.h
+++ b/drivers/gpu/drm/panfrost/panfrost_device.h
@@ -85,6 +85,7 @@ struct panfrost_device {
struct {
struct devfreq *devfreq;
+ struct opp_table *regulators_opp_table;
struct thermal_cooling_device *cooling;
ktime_t busy_time;
ktime_t idle_time;
--
2.24.1
On 07/01/2020 23:06, Martin Blumenstingl wrote:
> dev_pm_opp_set_rate() needs a reference to the regulator which should be
> updated when updating the GPU frequency. The name of the regulator has
> to be passed at initialization-time using dev_pm_opp_set_regulators().
> Add the call to dev_pm_opp_set_regulators() so dev_pm_opp_set_rate()
> will update the GPU regulator when updating the frequency (just like
> we did this manually before when we open-coded dev_pm_opp_set_rate()).
This patch causes a warning from debugfs on my firefly (RK3288) board:
debugfs: Directory 'ffa30000.gpu-mali' with parent 'vdd_gpu' already
present!
So it looks like the regulator is being added twice - but I haven't
investigated further.
> Fixes: 221bc77914cbcc ("drm/panfrost: Use generic code for devfreq")
> Reported-by: Robin Murphy <[email protected]>
> Signed-off-by: Martin Blumenstingl <[email protected]>
> ---
> drivers/gpu/drm/panfrost/panfrost_devfreq.c | 22 ++++++++++++++++++++-
> drivers/gpu/drm/panfrost/panfrost_device.h | 1 +
> 2 files changed, 22 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/panfrost/panfrost_devfreq.c b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
> index 170f6c8c9651..4f7999c7b44c 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_devfreq.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
> @@ -74,6 +74,7 @@ static struct devfreq_dev_profile panfrost_devfreq_profile = {
> int panfrost_devfreq_init(struct panfrost_device *pfdev)
> {
> int ret;
> + struct opp_table *opp_table;
> struct dev_pm_opp *opp;
> unsigned long cur_freq;
> struct device *dev = &pfdev->pdev->dev;
> @@ -84,9 +85,24 @@ int panfrost_devfreq_init(struct panfrost_device *pfdev)
> /* Optional, continue without devfreq */
> return 0;
>
> + opp_table = dev_pm_opp_set_regulators(dev,
> + (const char *[]){ "mali" },
> + 1);
> + if (IS_ERR(opp_table)) {
> + ret = PTR_ERR(opp_table);
> +
> + /* Continue if the optional regulator is missing */
> + if (ret != -ENODEV)
> + return ret;
> + } else {
> + pfdev->devfreq.regulators_opp_table = opp_table;
> + }
> +
> ret = dev_pm_opp_of_add_table(dev);
> - if (ret)
> + if (ret) {
> + dev_pm_opp_put_regulators(pfdev->devfreq.regulators_opp_table);
If we don't have a regulator then regulators_opp_table will be NULL and
sadly dev_pm_opp_put_regulators() doesn't handle a NULL argument. The
same applies to the two below calls obviously.
Steve
> return ret;
> + }
>
> panfrost_devfreq_reset(pfdev);
>
> @@ -95,6 +111,7 @@ int panfrost_devfreq_init(struct panfrost_device *pfdev)
> opp = devfreq_recommended_opp(dev, &cur_freq, 0);
> if (IS_ERR(opp)) {
> dev_pm_opp_of_remove_table(dev);
> + dev_pm_opp_put_regulators(pfdev->devfreq.regulators_opp_table);
> return PTR_ERR(opp);
> }
>
> @@ -106,6 +123,7 @@ int panfrost_devfreq_init(struct panfrost_device *pfdev)
> if (IS_ERR(devfreq)) {
> DRM_DEV_ERROR(dev, "Couldn't initialize GPU devfreq\n");
> dev_pm_opp_of_remove_table(dev);
> + dev_pm_opp_put_regulators(pfdev->devfreq.regulators_opp_table);
> return PTR_ERR(devfreq);
> }
> pfdev->devfreq.devfreq = devfreq;
> @@ -124,6 +142,8 @@ void panfrost_devfreq_fini(struct panfrost_device *pfdev)
> if (pfdev->devfreq.cooling)
> devfreq_cooling_unregister(pfdev->devfreq.cooling);
> dev_pm_opp_of_remove_table(&pfdev->pdev->dev);
> + if (pfdev->devfreq.regulators_opp_table)
> + dev_pm_opp_put_regulators(pfdev->devfreq.regulators_opp_table);
> }
>
> void panfrost_devfreq_resume(struct panfrost_device *pfdev)
> diff --git a/drivers/gpu/drm/panfrost/panfrost_device.h b/drivers/gpu/drm/panfrost/panfrost_device.h
> index 06713811b92c..4878b239e301 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_device.h
> +++ b/drivers/gpu/drm/panfrost/panfrost_device.h
> @@ -85,6 +85,7 @@ struct panfrost_device {
>
> struct {
> struct devfreq *devfreq;
> + struct opp_table *regulators_opp_table;
> struct thermal_cooling_device *cooling;
> ktime_t busy_time;
> ktime_t idle_time;
>
On 07/01/2020 23:06, Martin Blumenstingl wrote:
> If devfreq_recommended_opp() fails we need to undo
> dev_pm_opp_of_add_table() by calling dev_pm_opp_of_remove_table() (just
> like we do it in the other error-path below).
>
> Fixes: f3ba91228e8e91 ("drm/panfrost: Add initial panfrost driver")
> Signed-off-by: Martin Blumenstingl <[email protected]>
Reviewed-by: Steven Price <[email protected]>
> ---
> drivers/gpu/drm/panfrost/panfrost_devfreq.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/panfrost/panfrost_devfreq.c b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
> index 1471588763ce..170f6c8c9651 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_devfreq.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
> @@ -93,8 +93,10 @@ int panfrost_devfreq_init(struct panfrost_device *pfdev)
> cur_freq = clk_get_rate(pfdev->clock);
>
> opp = devfreq_recommended_opp(dev, &cur_freq, 0);
> - if (IS_ERR(opp))
> + if (IS_ERR(opp)) {
> + dev_pm_opp_of_remove_table(dev);
> return PTR_ERR(opp);
> + }
>
> panfrost_devfreq_profile.initial_freq = cur_freq;
> dev_pm_opp_put(opp);
>
On Thu, Jan 9, 2020 at 12:31 PM Steven Price <[email protected]> wrote:
>
> On 07/01/2020 23:06, Martin Blumenstingl wrote:
> > dev_pm_opp_set_rate() needs a reference to the regulator which should be
> > updated when updating the GPU frequency. The name of the regulator has
> > to be passed at initialization-time using dev_pm_opp_set_regulators().
> > Add the call to dev_pm_opp_set_regulators() so dev_pm_opp_set_rate()
> > will update the GPU regulator when updating the frequency (just like
> > we did this manually before when we open-coded dev_pm_opp_set_rate()).
>
> This patch causes a warning from debugfs on my firefly (RK3288) board:
>
> debugfs: Directory 'ffa30000.gpu-mali' with parent 'vdd_gpu' already
> present!
>
> So it looks like the regulator is being added twice - but I haven't
> investigated further.
I *think* it's because the regulator is already fetched by the
panfrost driver itself to enable it
(the devfreq code currently does not support enabling the regulator,
it can only control the voltage)
I'm not sure what to do about this though
[...]
> > ret = dev_pm_opp_of_add_table(dev);
> > - if (ret)
> > + if (ret) {
> > + dev_pm_opp_put_regulators(pfdev->devfreq.regulators_opp_table);
>
> If we don't have a regulator then regulators_opp_table will be NULL and
> sadly dev_pm_opp_put_regulators() doesn't handle a NULL argument. The
> same applies to the two below calls obviously.
good catch, thank you!
are you happy with the general approach here or do you think that
dev_pm_opp_set_regulators is the wrong way to go (for whatever
reason)?
Martin
On 09/01/2020 17:27, Martin Blumenstingl wrote:
> On Thu, Jan 9, 2020 at 12:31 PM Steven Price <[email protected]> wrote:
>>
>> On 07/01/2020 23:06, Martin Blumenstingl wrote:
>>> dev_pm_opp_set_rate() needs a reference to the regulator which should be
>>> updated when updating the GPU frequency. The name of the regulator has
>>> to be passed at initialization-time using dev_pm_opp_set_regulators().
>>> Add the call to dev_pm_opp_set_regulators() so dev_pm_opp_set_rate()
>>> will update the GPU regulator when updating the frequency (just like
>>> we did this manually before when we open-coded dev_pm_opp_set_rate()).
>>
>> This patch causes a warning from debugfs on my firefly (RK3288) board:
>>
>> debugfs: Directory 'ffa30000.gpu-mali' with parent 'vdd_gpu' already
>> present!
>>
>> So it looks like the regulator is being added twice - but I haven't
>> investigated further.
> I *think* it's because the regulator is already fetched by the
> panfrost driver itself to enable it
> (the devfreq code currently does not support enabling the regulator,
> it can only control the voltage)
>
> I'm not sure what to do about this though
Having a little play around with this, I think you can simply remove the
panfrost_regulator_init() call. This at least works for me - the call to
dev_pm_opp_set_regulators() seems to set everything up. However I
suspect you need to do this unconditionally even if there are no
operating points defined.
> [...]
>>> ret = dev_pm_opp_of_add_table(dev);
>>> - if (ret)
>>> + if (ret) {
>>> + dev_pm_opp_put_regulators(pfdev->devfreq.regulators_opp_table);
>>
>> If we don't have a regulator then regulators_opp_table will be NULL and
>> sadly dev_pm_opp_put_regulators() doesn't handle a NULL argument. The
>> same applies to the two below calls obviously.
> good catch, thank you!
> are you happy with the general approach here or do you think that
> dev_pm_opp_set_regulators is the wrong way to go (for whatever
> reason)?
To be honest this is an area I still don't fully understand. There's a
lot of magic helper functions and very little in the way of helpful
documentation to work out which are the right ones to call. It seems
reasonable to me, hopefully someone more in the know will chime in it
there's something fundamentally wrong!
Thanks,
Steve
Hi Steven,
On Mon, Jan 13, 2020 at 6:10 PM Steven Price <[email protected]> wrote:
>
> On 09/01/2020 17:27, Martin Blumenstingl wrote:
> > On Thu, Jan 9, 2020 at 12:31 PM Steven Price <[email protected]> wrote:
> >>
> >> On 07/01/2020 23:06, Martin Blumenstingl wrote:
> >>> dev_pm_opp_set_rate() needs a reference to the regulator which should be
> >>> updated when updating the GPU frequency. The name of the regulator has
> >>> to be passed at initialization-time using dev_pm_opp_set_regulators().
> >>> Add the call to dev_pm_opp_set_regulators() so dev_pm_opp_set_rate()
> >>> will update the GPU regulator when updating the frequency (just like
> >>> we did this manually before when we open-coded dev_pm_opp_set_rate()).
> >>
> >> This patch causes a warning from debugfs on my firefly (RK3288) board:
> >>
> >> debugfs: Directory 'ffa30000.gpu-mali' with parent 'vdd_gpu' already
> >> present!
> >>
> >> So it looks like the regulator is being added twice - but I haven't
> >> investigated further.
> > I *think* it's because the regulator is already fetched by the
> > panfrost driver itself to enable it
> > (the devfreq code currently does not support enabling the regulator,
> > it can only control the voltage)
> >
> > I'm not sure what to do about this though
>
> Having a little play around with this, I think you can simply remove the
> panfrost_regulator_init() call. This at least works for me - the call to
> dev_pm_opp_set_regulators() seems to set everything up. However I
> suspect you need to do this unconditionally even if there are no
> operating points defined.
I'm not sure if I can safely remove panfrost_regulator_init() because
it calls regulator_enable()
but there's no regulator_enable() equivalent in devfreq or OPP
I'm not sure how this is supposed to work
if someone has an idea: please let me know
> > [...]
> >>> ret = dev_pm_opp_of_add_table(dev);
> >>> - if (ret)
> >>> + if (ret) {
> >>> + dev_pm_opp_put_regulators(pfdev->devfreq.regulators_opp_table);
> >>
> >> If we don't have a regulator then regulators_opp_table will be NULL and
> >> sadly dev_pm_opp_put_regulators() doesn't handle a NULL argument. The
> >> same applies to the two below calls obviously.
> > good catch, thank you!
> > are you happy with the general approach here or do you think that
> > dev_pm_opp_set_regulators is the wrong way to go (for whatever
> > reason)?
>
> To be honest this is an area I still don't fully understand. There's a
> lot of magic helper functions and very little in the way of helpful
> documentation to work out which are the right ones to call. It seems
> reasonable to me, hopefully someone more in the know will chime in it
> there's something fundamentally wrong!
OK, if you know anybody who could help then please Cc them
Martin