Re: [PATCH 05/15] drm/panfrost: use spinlock instead of atomic
Hi Robin,
On Fri, 29 May 2020 at 14:20, Robin Murphy 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
> > ---
> [...]
> > 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
> > #include
> >
> > 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,
>
> 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.
Thanks for you review,
>
> Robin.
___
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel
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 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
---
[...]
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
#include
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.
___
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [PATCH 05/15] drm/panfrost: use spinlock instead of atomic
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.
Signed-off-by: Clément Péron
---
[...]
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
#include
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.
Given that, should this actually be considered a fix for 9e62b885f715
("drm/panfrost: Simplify devfreq utilisation tracking")?
Robin.
___
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [PATCH 05/15] drm/panfrost: use spinlock instead of atomic
On 10/05/2020 17:55, Clément Péron wrote:
Convert busy_count to a simple int protected by spinlock.
Signed-off-by: Clément Péron
Looks like a fairly mechanical cleanup.
Reviewed-by: Steven Price
---
drivers/gpu/drm/panfrost/panfrost_devfreq.c | 43 +++--
drivers/gpu/drm/panfrost/panfrost_devfreq.h | 10 -
2 files changed, 41 insertions(+), 12 deletions(-)
diff --git a/drivers/gpu/drm/panfrost/panfrost_devfreq.c
b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
index 962550363391..78753cfb59fb 100644
--- a/drivers/gpu/drm/panfrost/panfrost_devfreq.c
+++ b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
@@ -12,16 +12,12 @@
static void panfrost_devfreq_update_utilization(struct panfrost_devfreq *pfdevfreq)
{
- ktime_t now;
- ktime_t last;
-
- if (!pfdevfreq->devfreq)
- return;
+ ktime_t now, last;
now = ktime_get();
last = pfdevfreq->time_last_update;
- if (atomic_read(>busy_count) > 0)
+ if (pfdevfreq->busy_count > 0)
pfdevfreq->busy_time += ktime_sub(now, last);
else
pfdevfreq->idle_time += ktime_sub(now, last);
@@ -59,10 +55,14 @@ static int panfrost_devfreq_get_dev_status(struct device
*dev,
{
struct panfrost_device *pfdev = dev_get_drvdata(dev);
struct panfrost_devfreq *pfdevfreq = >pfdevfreq;
+ unsigned long irqflags;
+
+ status->current_frequency = clk_get_rate(pfdev->clock);
+
+ spin_lock_irqsave(>lock, irqflags);
panfrost_devfreq_update_utilization(pfdevfreq);
- status->current_frequency = clk_get_rate(pfdev->clock);
status->total_time = ktime_to_ns(ktime_add(pfdevfreq->busy_time,
pfdevfreq->idle_time));
@@ -70,6 +70,8 @@ static int panfrost_devfreq_get_dev_status(struct device *dev,
panfrost_devfreq_reset(pfdevfreq);
+ spin_unlock_irqrestore(>lock, irqflags);
+
dev_dbg(pfdev->dev, "busy %lu total %lu %lu %% freq %lu MHz\n",
status->busy_time, status->total_time,
status->busy_time / (status->total_time / 100),
@@ -100,6 +102,8 @@ int panfrost_devfreq_init(struct panfrost_device *pfdev)
else if (ret)
return ret;
+ spin_lock_init(>lock);
+
panfrost_devfreq_reset(pfdevfreq);
cur_freq = clk_get_rate(pfdev->clock);
@@ -162,15 +166,32 @@ void panfrost_devfreq_suspend(struct panfrost_device
*pfdev)
void panfrost_devfreq_record_busy(struct panfrost_devfreq *pfdevfreq)
{
+ unsigned long irqflags;
+
+ if (!pfdevfreq->devfreq)
+ return;
+
+ spin_lock_irqsave(>lock, irqflags);
+
panfrost_devfreq_update_utilization(pfdevfreq);
- atomic_inc(>busy_count);
+
+ pfdevfreq->busy_count++;
+
+ spin_unlock_irqrestore(>lock, irqflags);
}
void panfrost_devfreq_record_idle(struct panfrost_devfreq *pfdevfreq)
{
- int count;
+ unsigned long irqflags;
+
+ if (!pfdevfreq->devfreq)
+ return;
+
+ spin_lock_irqsave(>lock, irqflags);
panfrost_devfreq_update_utilization(pfdevfreq);
- count = atomic_dec_if_positive(>busy_count);
- WARN_ON(count < 0);
+
+ WARN_ON(--pfdevfreq->busy_count < 0);
+
+ spin_unlock_irqrestore(>lock, irqflags);
}
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
#include
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.
+*/
+ spinlock_t lock;
};
int panfrost_devfreq_init(struct panfrost_device *pfdev);
___
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel
