drm/exynos: fix vblank handling during dpms off

On 10/02/2014 08:52 PM, Andrzej Hajda wrote:
> Hi,
>
> +CC possible victims
>
> On 10/02/2014 12:52 PM, Inki Dae wrote:
>> On 2014년 10월 02일 17:58, Joonyoung Shim wrote:
>>> Hi Andrzej,
>>>
>>> On 10/01/2014 05:14 PM, Andrzej Hajda wrote:
>>>> The patch disables vblanks during dpms off only if pagefilp has
>>>> not been finished. It also replaces drm_vblank_off with drm_crtc_vblank_put.
>>>> It fixes issue with page_flip ioctl not being able to acquire vblank counter.
>>> This problem isn't related with pageflip, it just causes from
>>> 7ffd7a68511c710b84db3548a1997fd2625f580a commit (drm: Always reject
>>> drm_vblank_get() after drm_vblank_off()).
>>>
>>> We need to use drm_vblank_on() as a counterpart to drm_vblank_off()
>>> after the commit .
>
> This patch should break also other drivers, it seems at least following
> drms could be affected:
> armada, sti, tegra.

Indeed we (tegra) have just been hit by this. The problem seems to come 
from the fact that we have been using drm_vblank_pre_modeset, 
drm_vblank_post_modeset and drm_vblank_off conjointly. All these 
functions depend on the value of vblank->inmodeset, and 7ffd7a68511 
increases the vblank reference counter only in drm_vblank_off, which can 
result in the acquired reference never being released.

The following seems to fix this for Tegra, by stopping using 
drm_vblank_pre/post_modeset and relying on drm_vblank_off/on instead:


Thierry, does this look ok to you?

But there might be another issue, which is that calls to 
drm_vblank_get() will return -EINVAL if invoked between drm_blank_off 
and drm_blank_on. Is this really the desired behavior? Can it at least 
happen? If so, how are drivers supposed to react to this situation?
--
To unsubscribe from this list: send the line "unsubscribe linux-tegra" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

On Thu, Oct 09, 2014 at 02:43:02PM +0900, Alexandre Courbot wrote:
> But there might be another issue, which is that calls to  
> drm_vblank_get() will return -EINVAL if invoked between drm_blank_off  
> and drm_blank_on. Is this really the desired behavior? Can it at least  
> happen? If so, how are drivers supposed to react to this situation?

I've not yet seen the commit which causes this problem, but I hope
that drm_wait_vblank() isn't affected by this.  In current mainline,
drm_vblank_get() is used inside drm_wait_vblank(), which is called as
a result of userspace calling DRM_IOCTL_WAIT_VBLANK.

So, what is the effect of this change on user applications making use
of the vblank wait ioctl - and is that change intended?

On Thu, Oct 09, 2014 at 09:52:58AM +0100, Russell King - ARM Linux wrote:
> On Thu, Oct 09, 2014 at 02:43:02PM +0900, Alexandre Courbot wrote:
> > But there might be another issue, which is that calls to  
> > drm_vblank_get() will return -EINVAL if invoked between drm_blank_off  
> > and drm_blank_on. Is this really the desired behavior? Can it at least  
> > happen? If so, how are drivers supposed to react to this situation?
> 
> I've not yet seen the commit which causes this problem, but I hope
> that drm_wait_vblank() isn't affected by this.  In current mainline,
> drm_vblank_get() is used inside drm_wait_vblank(), which is called as
> a result of userspace calling DRM_IOCTL_WAIT_VBLANK.
> 
> So, what is the effect of this change on user applications making use
> of the vblank wait ioctl - and is that change intended?

There's no effect on user applications if the driver behaves properly.
As far as I can tell, every driver that calls drm_vblank_off() but not
drm_vblank_on() will break. You can easily test this by running libdrm
modetest -s ... -v, which instead of toggling between the test pattern
and an all-gray framebuffer will switch to the gray one once and then
hang.

I guess that was probably not intended, but according to the new rules
all these drivers have now become buggy. So before merging this patch I
think we need to fix existing drivers to avoid regressions.

Thierry

On Thu, Oct 09, 2014 at 02:43:02PM +0900, Alexandre Courbot wrote:
> On 10/02/2014 08:52 PM, Andrzej Hajda wrote:
> >Hi,
> >
> >+CC possible victims
> >
> >On 10/02/2014 12:52 PM, Inki Dae wrote:
> >>On 2014년 10월 02일 17:58, Joonyoung Shim wrote:
> >>>Hi Andrzej,
> >>>
> >>>On 10/01/2014 05:14 PM, Andrzej Hajda wrote:
> >>>>The patch disables vblanks during dpms off only if pagefilp has
> >>>>not been finished. It also replaces drm_vblank_off with drm_crtc_vblank_put.
> >>>>It fixes issue with page_flip ioctl not being able to acquire vblank counter.
> >>>This problem isn't related with pageflip, it just causes from
> >>>7ffd7a68511c710b84db3548a1997fd2625f580a commit (drm: Always reject
> >>>drm_vblank_get() after drm_vblank_off()).
> >>>
> >>>We need to use drm_vblank_on() as a counterpart to drm_vblank_off()
> >>>after the commit .
> >
> >This patch should break also other drivers, it seems at least following
> >drms could be affected:
> >armada, sti, tegra.
> 
> Indeed we (tegra) have just been hit by this. The problem seems to come from
> the fact that we have been using drm_vblank_pre_modeset,
> drm_vblank_post_modeset and drm_vblank_off conjointly. All these functions
> depend on the value of vblank->inmodeset, and 7ffd7a68511 increases the
> vblank reference counter only in drm_vblank_off, which can result in the
> acquired reference never being released.
> 
> The following seems to fix this for Tegra, by stopping using
> drm_vblank_pre/post_modeset and relying on drm_vblank_off/on instead:
> 
> diff --git a/drivers/gpu/drm/tegra/dc.c b/drivers/gpu/drm/tegra/dc.c
> index b08df07cad47..3955d81236d0 100644
> --- a/drivers/gpu/drm/tegra/dc.c
> +++ b/drivers/gpu/drm/tegra/dc.c
> @@ -739,7 +739,6 @@ static const struct drm_crtc_funcs tegra_crtc_funcs = {
> 
>  static void tegra_crtc_disable(struct drm_crtc *crtc)
>  {
> -       struct tegra_dc *dc = to_tegra_dc(crtc);
>         struct drm_device *drm = crtc->dev;
>         struct drm_plane *plane;
> 
> @@ -755,7 +754,7 @@ static void tegra_crtc_disable(struct drm_crtc *crtc)
>                 }
>         }
> 
> -       drm_vblank_off(drm, dc->pipe);
> +       drm_crtc_vblank_off(crtc);
>  }
> 
>  static bool tegra_crtc_mode_fixup(struct drm_crtc *crtc,
> @@ -844,7 +843,7 @@ static int tegra_crtc_mode_set(struct drm_crtc *crtc,
>         u32 value;
>         int err;
> 
> -       drm_vblank_pre_modeset(crtc->dev, dc->pipe);
> +       drm_crtc_vblank_off(crtc);
> 
>         err = tegra_crtc_setup_clk(crtc, mode);
>         if (err) {
> @@ -946,7 +945,7 @@ static void tegra_crtc_commit(struct drm_crtc *crtc)
>         value = GENERAL_ACT_REQ | WIN_A_ACT_REQ;
>         tegra_dc_writel(dc, value, DC_CMD_STATE_CONTROL);
> 
> -       drm_vblank_post_modeset(crtc->dev, dc->pipe);
> +       drm_crtc_vblank_on(crtc);
>  }
> 
>  static void tegra_crtc_load_lut(struct drm_crtc *crtc)
> 
> Thierry, does this look ok to you?

Yes, that looks like almost the same patch that I sent out yesterday.
The difference is that I didn't replace the drm_vblank_pre_modeset()
call with drm_vblank_off() like you did, but rather just dropped the
former.

I /think/ your version is more correct in that regard.

Thierry

> But there might be another issue, which is that calls to drm_vblank_get()
> will return -EINVAL if invoked between drm_blank_off and drm_blank_on. Is
> this really the desired behavior? Can it at least happen? If so, how are
> drivers supposed to react to this situation?

It shouldn't happen. If drm_vblank_off() and drm_vblank_on() are called
around a modeset they should never conflict with drm_vblank_get(), at
least on Tegra, because the modeset and page-flip IOCTLs will be
serialized.

Thierry

On Thu, Oct 9, 2014 at 7:08 PM, Thierry Reding <treding@nvidia.com> wrote:
> On Thu, Oct 09, 2014 at 02:43:02PM +0900, Alexandre Courbot wrote:
>> On 10/02/2014 08:52 PM, Andrzej Hajda wrote:
>> >Hi,
>> >
>> >+CC possible victims
>> >
>> >On 10/02/2014 12:52 PM, Inki Dae wrote:
>> >>On 2014년 10월 02일 17:58, Joonyoung Shim wrote:
>> >>>Hi Andrzej,
>> >>>
>> >>>On 10/01/2014 05:14 PM, Andrzej Hajda wrote:
>> >>>>The patch disables vblanks during dpms off only if pagefilp has
>> >>>>not been finished. It also replaces drm_vblank_off with drm_crtc_vblank_put.
>> >>>>It fixes issue with page_flip ioctl not being able to acquire vblank counter.
>> >>>This problem isn't related with pageflip, it just causes from
>> >>>7ffd7a68511c710b84db3548a1997fd2625f580a commit (drm: Always reject
>> >>>drm_vblank_get() after drm_vblank_off()).
>> >>>
>> >>>We need to use drm_vblank_on() as a counterpart to drm_vblank_off()
>> >>>after the commit .
>> >
>> >This patch should break also other drivers, it seems at least following
>> >drms could be affected:
>> >armada, sti, tegra.
>>
>> Indeed we (tegra) have just been hit by this. The problem seems to come from
>> the fact that we have been using drm_vblank_pre_modeset,
>> drm_vblank_post_modeset and drm_vblank_off conjointly. All these functions
>> depend on the value of vblank->inmodeset, and 7ffd7a68511 increases the
>> vblank reference counter only in drm_vblank_off, which can result in the
>> acquired reference never being released.
>>
>> The following seems to fix this for Tegra, by stopping using
>> drm_vblank_pre/post_modeset and relying on drm_vblank_off/on instead:
>>
>> diff --git a/drivers/gpu/drm/tegra/dc.c b/drivers/gpu/drm/tegra/dc.c
>> index b08df07cad47..3955d81236d0 100644
>> --- a/drivers/gpu/drm/tegra/dc.c
>> +++ b/drivers/gpu/drm/tegra/dc.c
>> @@ -739,7 +739,6 @@ static const struct drm_crtc_funcs tegra_crtc_funcs = {
>>
>>  static void tegra_crtc_disable(struct drm_crtc *crtc)
>>  {
>> -       struct tegra_dc *dc = to_tegra_dc(crtc);
>>         struct drm_device *drm = crtc->dev;
>>         struct drm_plane *plane;
>>
>> @@ -755,7 +754,7 @@ static void tegra_crtc_disable(struct drm_crtc *crtc)
>>                 }
>>         }
>>
>> -       drm_vblank_off(drm, dc->pipe);
>> +       drm_crtc_vblank_off(crtc);
>>  }
>>
>>  static bool tegra_crtc_mode_fixup(struct drm_crtc *crtc,
>> @@ -844,7 +843,7 @@ static int tegra_crtc_mode_set(struct drm_crtc *crtc,
>>         u32 value;
>>         int err;
>>
>> -       drm_vblank_pre_modeset(crtc->dev, dc->pipe);
>> +       drm_crtc_vblank_off(crtc);
>>
>>         err = tegra_crtc_setup_clk(crtc, mode);
>>         if (err) {
>> @@ -946,7 +945,7 @@ static void tegra_crtc_commit(struct drm_crtc *crtc)
>>         value = GENERAL_ACT_REQ | WIN_A_ACT_REQ;
>>         tegra_dc_writel(dc, value, DC_CMD_STATE_CONTROL);
>>
>> -       drm_vblank_post_modeset(crtc->dev, dc->pipe);
>> +       drm_crtc_vblank_on(crtc);
>>  }
>>
>>  static void tegra_crtc_load_lut(struct drm_crtc *crtc)
>>
>> Thierry, does this look ok to you?
>
> Yes, that looks like almost the same patch that I sent out yesterday.
> The difference is that I didn't replace the drm_vblank_pre_modeset()
> call with drm_vblank_off() like you did, but rather just dropped the
> former.
>
> I /think/ your version is more correct in that regard.

Feel free to take that extra line in your patch then. ;)

>
> Thierry
>
>> But there might be another issue, which is that calls to drm_vblank_get()
>> will return -EINVAL if invoked between drm_blank_off and drm_blank_on. Is
>> this really the desired behavior? Can it at least happen? If so, how are
>> drivers supposed to react to this situation?
>
> It shouldn't happen. If drm_vblank_off() and drm_vblank_on() are called
> around a modeset they should never conflict with drm_vblank_get(), at
> least on Tegra, because the modeset and page-flip IOCTLs will be
> serialized.

Ok, that's good. I was just wondering whether this case has been thought of.

Actually, and seeing how drm_vblank_pre/post_modeset() have become
useless (if not harmful), maybe it would be a good idea to come with a
fixup patch that gets rid of them altogether and make their callers
invoke drm_vblank_off/on() instead?
--
To unsubscribe from this list: send the line "unsubscribe linux-tegra" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html