Nouveau - Aug 2019 - [PATCH 1/2] drm/nouveau/dispnv04: Grab/put runtime PM refs on DPMS on/off

On Wed, Aug 07, 2019 at 05:33:00PM -0400, Lyude Paul
wrote:
> The code claims to grab a runtime PM ref when at least one CRTC is
> active, but that's not actually the case as we grab a runtime PM ref
> whenever a CRTC is enabled regardless of it's DPMS state. Meaning that
> we can end up keeping the GPU awake when there are no screens enabled,
> something we don't really want to do.
> 
> Note that we fixed this same issue for nv50 a while ago in:
> 
> commit e5d54f193572 ("drm/nouveau/drm/nouveau: Fix runtime PM leak in
> nv50_disp_atomic_commit()")
> 
> Since we're about to remove nouveau_drm->have_disp_power_ref in the
next
> commit, let's also simplify the RPM code here while we're at it:
grab a
> ref during a modeset, grab additional RPM refs for each CRTC enabled by
> said modeset, and drop an RPM ref for each CRTC disabled by said
> modeset. This allows us to keep the GPU awake whenever screens are
> turned on, without needing to use nouveau_drm->have_disp_power_ref.
> 
> Signed-off-by: Lyude Paul <lyude at redhat.com>
> ---
>  drivers/gpu/drm/nouveau/dispnv04/crtc.c | 18 ++++--------------
>  1 file changed, 4 insertions(+), 14 deletions(-)
> 
> diff --git a/drivers/gpu/drm/nouveau/dispnv04/crtc.c
b/drivers/gpu/drm/nouveau/dispnv04/crtc.c
> index f22f01020625..08ad8e3b9cd2 100644
> --- a/drivers/gpu/drm/nouveau/dispnv04/crtc.c
> +++ b/drivers/gpu/drm/nouveau/dispnv04/crtc.c
> @@ -183,6 +183,10 @@ nv_crtc_dpms(struct drm_crtc *crtc, int mode)
>  		return;
>  
>  	nv_crtc->last_dpms = mode;
> +	if (mode == DRM_MODE_DPMS_ON)
> +		pm_runtime_get_noresume(dev->dev);
> +	else
> +		pm_runtime_put_noidle(dev->dev);
it's after we filter out duplicate operations, so that part looks good.
But not all of nouveau's legacy helper crtc callbacks go throuh ->dpms I
think: nv_crtc_disable doesn't, and crtc helpers use that in preference
over ->dpms in some cases.

I think the only way to actually hit that path is if you switch an active
connector from an active CRTC to an inactive one. This implicitly disables
the crtc (well, should, nv_crtc_disable doesn't seem to shut down hw), and
I think would leak your runtime PM reference here. At least temporarily.

No idea how to best fix that. Aside from "use atomic" :-)

Cheers, Daniel
>  
>  	if (nv_two_heads(dev))
>  		NVSetOwner(dev, nv_crtc->index);
> @@ -1045,7 +1049,6 @@ nouveau_crtc_set_config(struct drm_mode_set *set,
>  
>  	dev = set->crtc->dev;
>  
> -	/* get a pm reference here */
>  	ret = pm_runtime_get_sync(dev->dev);
>  	if (ret < 0 && ret != -EACCES)
>  		return ret;
> @@ -1061,19 +1064,6 @@ nouveau_crtc_set_config(struct drm_mode_set *set,
>  	}
>  
>  	pm_runtime_mark_last_busy(dev->dev);
> -	/* if we have active crtcs and we don't have a power ref,
> -	   take the current one */
> -	if (active && !drm->have_disp_power_ref) {
> -		drm->have_disp_power_ref = true;
> -		return ret;
> -	}
> -	/* if we have no active crtcs, then drop the power ref
> -	   we got before */
> -	if (!active && drm->have_disp_power_ref) {
> -		pm_runtime_put_autosuspend(dev->dev);
> -		drm->have_disp_power_ref = false;
> -	}
> -	/* drop the power reference we got coming in here */
>  	pm_runtime_put_autosuspend(dev->dev);
>  	return ret;
>  }
> -- 
> 2.21.0
> 
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

On Wed, Aug 7, 2019 at 5:55 PM Daniel Vetter <daniel at ffwll.ch>
wrote:
>
> On Wed, Aug 07, 2019 at 05:33:00PM -0400, Lyude Paul wrote:
> > The code claims to grab a runtime PM ref when at least one CRTC is
> > active, but that's not actually the case as we grab a runtime PM
ref
> > whenever a CRTC is enabled regardless of it's DPMS state. Meaning
that
> > we can end up keeping the GPU awake when there are no screens enabled,
> > something we don't really want to do.
> >
> > Note that we fixed this same issue for nv50 a while ago in:
> >
> > commit e5d54f193572 ("drm/nouveau/drm/nouveau: Fix runtime PM
leak in
> > nv50_disp_atomic_commit()")
> >
> > Since we're about to remove nouveau_drm->have_disp_power_ref in
the next
> > commit, let's also simplify the RPM code here while we're at
it: grab a
> > ref during a modeset, grab additional RPM refs for each CRTC enabled
by
> > said modeset, and drop an RPM ref for each CRTC disabled by said
> > modeset. This allows us to keep the GPU awake whenever screens are
> > turned on, without needing to use nouveau_drm->have_disp_power_ref.
> >
> > Signed-off-by: Lyude Paul <lyude at redhat.com>
> > ---
> >  drivers/gpu/drm/nouveau/dispnv04/crtc.c | 18 ++++--------------
> >  1 file changed, 4 insertions(+), 14 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/nouveau/dispnv04/crtc.c
b/drivers/gpu/drm/nouveau/dispnv04/crtc.c
> > index f22f01020625..08ad8e3b9cd2 100644
> > --- a/drivers/gpu/drm/nouveau/dispnv04/crtc.c
> > +++ b/drivers/gpu/drm/nouveau/dispnv04/crtc.c
> > @@ -183,6 +183,10 @@ nv_crtc_dpms(struct drm_crtc *crtc, int mode)
> >               return;
> >
> >       nv_crtc->last_dpms = mode;
> > +     if (mode == DRM_MODE_DPMS_ON)
> > +             pm_runtime_get_noresume(dev->dev);
> > +     else
> > +             pm_runtime_put_noidle(dev->dev);
>
> it's after we filter out duplicate operations, so that part looks good.
> But not all of nouveau's legacy helper crtc callbacks go throuh
->dpms I
> think: nv_crtc_disable doesn't, and crtc helpers use that in preference
> over ->dpms in some cases.
>
> I think the only way to actually hit that path is if you switch an active
> connector from an active CRTC to an inactive one. This implicitly disables
> the crtc (well, should, nv_crtc_disable doesn't seem to shut down hw),
and
> I think would leak your runtime PM reference here. At least temporarily.
>
> No idea how to best fix that. Aside from "use atomic" :-)
Not sure if this is relevant to the discussion at hand, but I'd like
to point out that dispnv04 is for pre-nv50 things, which definitely
didn't support any kind of ACPI-based runtime suspend.

  -ilia

On Wed, 2019-08-07 at 19:06 -0400, Ilia Mirkin wrote:
> On Wed, Aug 7, 2019 at 5:55 PM Daniel Vetter <daniel at ffwll.ch>
wrote:
> > On Wed, Aug 07, 2019 at 05:33:00PM -0400, Lyude Paul wrote:
> > > The code claims to grab a runtime PM ref when at least one CRTC
is
> > > active, but that's not actually the case as we grab a runtime
PM ref
> > > whenever a CRTC is enabled regardless of it's DPMS state.
Meaning that
> > > we can end up keeping the GPU awake when there are no screens
enabled,
> > > something we don't really want to do.
> > > 
> > > Note that we fixed this same issue for nv50 a while ago in:
> > > 
> > > commit e5d54f193572 ("drm/nouveau/drm/nouveau: Fix runtime
PM leak in
> > > nv50_disp_atomic_commit()")
> > > 
> > > Since we're about to remove
nouveau_drm->have_disp_power_ref in the next
> > > commit, let's also simplify the RPM code here while we're
at it: grab a
> > > ref during a modeset, grab additional RPM refs for each CRTC
enabled by
> > > said modeset, and drop an RPM ref for each CRTC disabled by said
> > > modeset. This allows us to keep the GPU awake whenever screens
are
> > > turned on, without needing to use
nouveau_drm->have_disp_power_ref.
> > > 
> > > Signed-off-by: Lyude Paul <lyude at redhat.com>
> > > ---
> > >  drivers/gpu/drm/nouveau/dispnv04/crtc.c | 18 ++++--------------
> > >  1 file changed, 4 insertions(+), 14 deletions(-)
> > > 
> > > diff --git a/drivers/gpu/drm/nouveau/dispnv04/crtc.c
> > > b/drivers/gpu/drm/nouveau/dispnv04/crtc.c
> > > index f22f01020625..08ad8e3b9cd2 100644
> > > --- a/drivers/gpu/drm/nouveau/dispnv04/crtc.c
> > > +++ b/drivers/gpu/drm/nouveau/dispnv04/crtc.c
> > > @@ -183,6 +183,10 @@ nv_crtc_dpms(struct drm_crtc *crtc, int
mode)
> > >               return;
> > > 
> > >       nv_crtc->last_dpms = mode;
> > > +     if (mode == DRM_MODE_DPMS_ON)
> > > +             pm_runtime_get_noresume(dev->dev);
> > > +     else
> > > +             pm_runtime_put_noidle(dev->dev);
> > 
> > it's after we filter out duplicate operations, so that part looks
good.
> > But not all of nouveau's legacy helper crtc callbacks go throuh
->dpms I
> > think: nv_crtc_disable doesn't, and crtc helpers use that in
preference
> > over ->dpms in some cases.
> > 
> > I think the only way to actually hit that path is if you switch an
active
> > connector from an active CRTC to an inactive one. This implicitly
disables
> > the crtc (well, should, nv_crtc_disable doesn't seem to shut down
hw), and
> > I think would leak your runtime PM reference here. At least
temporarily.
> > 
> > No idea how to best fix that. Aside from "use atomic" :-)
> 
> Not sure if this is relevant to the discussion at hand, but I'd like
> to point out that dispnv04 is for pre-nv50 things, which definitely
> didn't support any kind of ACPI-based runtime suspend.
I thought it was really suspicious that such an old chipset had any kind of
runtime PM, but didn't question the code lol. I guess a more appropriate
patch
would be to just remove runtime PM support entirely for pre-nv50. Will respin
soon and do this.
> 
>   -ilia