diff mbox

drm/i915: flush delayed_resume_work when suspending

Message ID 1403905911-3859-1-git-send-email-przanoni@gmail.com (mailing list archive)
State New, archived
Headers show

Commit Message

Paulo Zanoni June 27, 2014, 9:51 p.m. UTC
From: Paulo Zanoni <paulo.r.zanoni@intel.com>

It is possible that, by the time we run i915_drm_freeze(),
delayed_resume_work was already queued but did not run yet. If it
still didn't run after intel_runtime_pm_disable_interrupts(), by the
time it runs it will try to change the interrupt registers with the
interrupts already disabled, which will trigger a WARN. We can
reliably reproduce this with the pm_rpm system-suspend test case.

In order to avoid the problem, we have to flush the work before
disabling the interrupts. We could also cancel the work instead of
flushing it, but that would require us to put a runtime PM reference -
and any other resource we may need in the future - in case the work
was already queued, so I believe flushing the work is more
future-proof, although less efficient. But I can also change this part
if someone requests.

Another thing I tried was to move the intel_suspend_gt_powersave()
call to before intel_runtime_pm_disable_interrupts(), but since that
function needs to be called after the interrupts are already disabled,
due to dev_priv->rps.work, this strategy didn't work.

Testcase: igt/pm_rpm/system-suspend
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=80517
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
---
 drivers/gpu/drm/i915/i915_drv.c | 2 ++
 1 file changed, 2 insertions(+)

Comments

Rodrigo Vivi June 27, 2014, 10:30 p.m. UTC | #1
I have the feeling the safest side would be disable rc6 on resume instead
of force its enabling... or am I missing something?
why don't you just cancel the work? and put another after resume?

but if the patch really solves the problem and this is what you meant feel
free to use:
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>



On Fri, Jun 27, 2014 at 2:51 PM, Paulo Zanoni <przanoni@gmail.com> wrote:

> From: Paulo Zanoni <paulo.r.zanoni@intel.com>
>
> It is possible that, by the time we run i915_drm_freeze(),
> delayed_resume_work was already queued but did not run yet. If it
> still didn't run after intel_runtime_pm_disable_interrupts(), by the
> time it runs it will try to change the interrupt registers with the
> interrupts already disabled, which will trigger a WARN. We can
> reliably reproduce this with the pm_rpm system-suspend test case.
>
> In order to avoid the problem, we have to flush the work before
> disabling the interrupts. We could also cancel the work instead of
> flushing it, but that would require us to put a runtime PM reference -
> and any other resource we may need in the future - in case the work
> was already queued, so I believe flushing the work is more
> future-proof, although less efficient. But I can also change this part
> if someone requests.
>
> Another thing I tried was to move the intel_suspend_gt_powersave()
> call to before intel_runtime_pm_disable_interrupts(), but since that
> function needs to be called after the interrupts are already disabled,
> due to dev_priv->rps.work, this strategy didn't work.
>
> Testcase: igt/pm_rpm/system-suspend
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=80517
> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
> ---
>  drivers/gpu/drm/i915/i915_drv.c | 2 ++
>  1 file changed, 2 insertions(+)
>
> diff --git a/drivers/gpu/drm/i915/i915_drv.c
> b/drivers/gpu/drm/i915/i915_drv.c
> index e64547e..672694b 100644
> --- a/drivers/gpu/drm/i915/i915_drv.c
> +++ b/drivers/gpu/drm/i915/i915_drv.c
> @@ -524,6 +524,8 @@ static int i915_drm_freeze(struct drm_device *dev)
>                         return error;
>                 }
>
> +               flush_delayed_work(&dev_priv->rps.delayed_resume_work);
> +
>                 intel_runtime_pm_disable_interrupts(dev);
>                 dev_priv->enable_hotplug_processing = false;
>
> --
> 2.0.0
>
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx
>
Paulo Zanoni June 30, 2014, 7:22 p.m. UTC | #2
2014-06-27 19:30 GMT-03:00 Rodrigo Vivi <rodrigo.vivi@gmail.com>:
> I have the feeling the safest side would be disable rc6 on resume instead of
> force its enabling... or am I missing something?

It will be enabled, then disabled.

> why don't you just cancel the work? and put another after resume?
>
> but if the patch really solves the problem and this is what you meant feel
> free to use:

What you're suggesting is the "We could also" case mentioned in the
second paragraph of the commit message. I even wrote and tested that
patch, but Jesse seemed to prefer the "flush" version instead of the
"cancel" one. I'll send the other version to the list, then reviewers
and maintainers can decide which one they prefer :)

> Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
>
>
>
> On Fri, Jun 27, 2014 at 2:51 PM, Paulo Zanoni <przanoni@gmail.com> wrote:
>>
>> From: Paulo Zanoni <paulo.r.zanoni@intel.com>
>>
>> It is possible that, by the time we run i915_drm_freeze(),
>> delayed_resume_work was already queued but did not run yet. If it
>> still didn't run after intel_runtime_pm_disable_interrupts(), by the
>> time it runs it will try to change the interrupt registers with the
>> interrupts already disabled, which will trigger a WARN. We can
>> reliably reproduce this with the pm_rpm system-suspend test case.
>>
>> In order to avoid the problem, we have to flush the work before
>> disabling the interrupts. We could also cancel the work instead of
>> flushing it, but that would require us to put a runtime PM reference -
>> and any other resource we may need in the future - in case the work
>> was already queued, so I believe flushing the work is more
>> future-proof, although less efficient. But I can also change this part
>> if someone requests.
>>
>> Another thing I tried was to move the intel_suspend_gt_powersave()
>> call to before intel_runtime_pm_disable_interrupts(), but since that
>> function needs to be called after the interrupts are already disabled,
>> due to dev_priv->rps.work, this strategy didn't work.
>>
>> Testcase: igt/pm_rpm/system-suspend
>> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=80517
>> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
>> ---
>>  drivers/gpu/drm/i915/i915_drv.c | 2 ++
>>  1 file changed, 2 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/i915/i915_drv.c
>> b/drivers/gpu/drm/i915/i915_drv.c
>> index e64547e..672694b 100644
>> --- a/drivers/gpu/drm/i915/i915_drv.c
>> +++ b/drivers/gpu/drm/i915/i915_drv.c
>> @@ -524,6 +524,8 @@ static int i915_drm_freeze(struct drm_device *dev)
>>                         return error;
>>                 }
>>
>> +               flush_delayed_work(&dev_priv->rps.delayed_resume_work);
>> +
>>                 intel_runtime_pm_disable_interrupts(dev);
>>                 dev_priv->enable_hotplug_processing = false;
>>
>> --
>> 2.0.0
>>
>> _______________________________________________
>> Intel-gfx mailing list
>> Intel-gfx@lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/intel-gfx
>
>
>
>
> --
> Rodrigo Vivi
> Blog: http://blog.vivi.eng.br
>
Rodrigo Vivi June 30, 2014, 8:10 p.m. UTC | #3
On Mon, Jun 30, 2014 at 12:22 PM, Paulo Zanoni <przanoni@gmail.com> wrote:

> 2014-06-27 19:30 GMT-03:00 Rodrigo Vivi <rodrigo.vivi@gmail.com>:
> > I have the feeling the safest side would be disable rc6 on resume
> instead of
> > force its enabling... or am I missing something?
>
> It will be enabled, then disabled.
>

oh that's true!


>
> > why don't you just cancel the work? and put another after resume?
> >
> > but if the patch really solves the problem and this is what you meant
> feel
> > free to use:
>
> What you're suggesting is the "We could also" case mentioned in the
> second paragraph of the commit message. I even wrote and tested that
> patch,


Yeah, reading again this is exactly what I had in mind.


> but Jesse seemed to prefer the "flush" version instead of the
> "cancel" one. I'll send the other version to the list, then reviewers
> and maintainers can decide which one they prefer :)
>

I don't have stronger preferences. So, feel free to use:
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>

Thanks for the explanations,
Rodrigo.


>
> > Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
> >
> >
> >
> > On Fri, Jun 27, 2014 at 2:51 PM, Paulo Zanoni <przanoni@gmail.com>
> wrote:
> >>
> >> From: Paulo Zanoni <paulo.r.zanoni@intel.com>
> >>
> >> It is possible that, by the time we run i915_drm_freeze(),
> >> delayed_resume_work was already queued but did not run yet. If it
> >> still didn't run after intel_runtime_pm_disable_interrupts(), by the
> >> time it runs it will try to change the interrupt registers with the
> >> interrupts already disabled, which will trigger a WARN. We can
> >> reliably reproduce this with the pm_rpm system-suspend test case.
> >>
> >> In order to avoid the problem, we have to flush the work before
> >> disabling the interrupts. We could also cancel the work instead of
> >> flushing it, but that would require us to put a runtime PM reference -
> >> and any other resource we may need in the future - in case the work
> >> was already queued, so I believe flushing the work is more
> >> future-proof, although less efficient. But I can also change this part
> >> if someone requests.
> >>
> >> Another thing I tried was to move the intel_suspend_gt_powersave()
> >> call to before intel_runtime_pm_disable_interrupts(), but since that
> >> function needs to be called after the interrupts are already disabled,
> >> due to dev_priv->rps.work, this strategy didn't work.
> >>
> >> Testcase: igt/pm_rpm/system-suspend
> >> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=80517
> >> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
> >> ---
> >>  drivers/gpu/drm/i915/i915_drv.c | 2 ++
> >>  1 file changed, 2 insertions(+)
> >>
> >> diff --git a/drivers/gpu/drm/i915/i915_drv.c
> >> b/drivers/gpu/drm/i915/i915_drv.c
> >> index e64547e..672694b 100644
> >> --- a/drivers/gpu/drm/i915/i915_drv.c
> >> +++ b/drivers/gpu/drm/i915/i915_drv.c
> >> @@ -524,6 +524,8 @@ static int i915_drm_freeze(struct drm_device *dev)
> >>                         return error;
> >>                 }
> >>
> >> +               flush_delayed_work(&dev_priv->rps.delayed_resume_work);
> >> +
> >>                 intel_runtime_pm_disable_interrupts(dev);
> >>                 dev_priv->enable_hotplug_processing = false;
> >>
> >> --
> >> 2.0.0
> >>
> >> _______________________________________________
> >> Intel-gfx mailing list
> >> Intel-gfx@lists.freedesktop.org
> >> http://lists.freedesktop.org/mailman/listinfo/intel-gfx
> >
> >
> >
> >
> > --
> > Rodrigo Vivi
> > Blog: http://blog.vivi.eng.br
> >
>
>
>
> --
> Paulo Zanoni
>
Jani Nikula July 1, 2014, 2:49 p.m. UTC | #4
On Sat, 28 Jun 2014, Paulo Zanoni <przanoni@gmail.com> wrote:
> From: Paulo Zanoni <paulo.r.zanoni@intel.com>
>
> It is possible that, by the time we run i915_drm_freeze(),
> delayed_resume_work was already queued but did not run yet. If it
> still didn't run after intel_runtime_pm_disable_interrupts(), by the
> time it runs it will try to change the interrupt registers with the
> interrupts already disabled, which will trigger a WARN. We can
> reliably reproduce this with the pm_rpm system-suspend test case.
>
> In order to avoid the problem, we have to flush the work before
> disabling the interrupts. We could also cancel the work instead of
> flushing it, but that would require us to put a runtime PM reference -
> and any other resource we may need in the future - in case the work
> was already queued, so I believe flushing the work is more
> future-proof, although less efficient. But I can also change this part
> if someone requests.
>
> Another thing I tried was to move the intel_suspend_gt_powersave()
> call to before intel_runtime_pm_disable_interrupts(), but since that
> function needs to be called after the interrupts are already disabled,
> due to dev_priv->rps.work, this strategy didn't work.
>
> Testcase: igt/pm_rpm/system-suspend
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=80517
> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>

Pushed to dinq, thanks for the patch and review.

BR,
Jani.


> ---
>  drivers/gpu/drm/i915/i915_drv.c | 2 ++
>  1 file changed, 2 insertions(+)
>
> diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
> index e64547e..672694b 100644
> --- a/drivers/gpu/drm/i915/i915_drv.c
> +++ b/drivers/gpu/drm/i915/i915_drv.c
> @@ -524,6 +524,8 @@ static int i915_drm_freeze(struct drm_device *dev)
>  			return error;
>  		}
>  
> +		flush_delayed_work(&dev_priv->rps.delayed_resume_work);
> +
>  		intel_runtime_pm_disable_interrupts(dev);
>  		dev_priv->enable_hotplug_processing = false;
>  
> -- 
> 2.0.0
>
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx
diff mbox

Patch

diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
index e64547e..672694b 100644
--- a/drivers/gpu/drm/i915/i915_drv.c
+++ b/drivers/gpu/drm/i915/i915_drv.c
@@ -524,6 +524,8 @@  static int i915_drm_freeze(struct drm_device *dev)
 			return error;
 		}
 
+		flush_delayed_work(&dev_priv->rps.delayed_resume_work);
+
 		intel_runtime_pm_disable_interrupts(dev);
 		dev_priv->enable_hotplug_processing = false;