On Wed, Apr 08, 2020 at 07:18:11PM +0300, Lisovskiy, Stanislav wrote:
> On Wed, Apr 08, 2020 at 06:54:09PM +0300, Lisovskiy, Stanislav wrote:
> > On Wed, Apr 08, 2020 at 05:55:02PM +0300, Ville Syrjälä wrote:
> > > On Wed, Apr 08, 2020 at 10:58:04AM +0300, Lisovskiy, Stanislav wrote:
> > > > On Tue, Apr 07, 2020 at 10:01:28PM +0300, Ville Syrjälä wrote:
> > > > > On Fri, Apr 03, 2020 at 09:20:03AM +0300, Stanislav Lisovskiy wrote:
> > > > > > Addressing one of the comments, recommending to extract platform
> > > > > > specific code from intel_can_enable_sagv as a preparation, before
> > > > > > we are going to add support for tgl+.
> > > > > >
> > > > > > Current code in intel_can_enable_sagv is valid only for skl,
> > > > > > so this patch adds also proper support for icl, subsequent
> > > > > > patches will add support for tgl+, combined with other required
> > > > > > changes.
> > > > > >
> > > > > > v2: - Renamed icl_can_enable_sagv into icl_crtc_can_enable_sagv(Ville)
> > > > > > - Removed dev variables(Ville)
> > > > > > - Constified crtc/plane_state in icl_crtc_can_enable_sagv
> > > > > > function(Ville)
> > > > > > - Added hw.active check(Ville)
> > > > > > - Refactored if ladder(Ville)
> > > > > >
> > > > > > Signed-off-by: Stanislav Lisovskiy <stanislav.lisovskiy@xxxxxxxxx>
> > > > > > ---
> > > > > > drivers/gpu/drm/i915/intel_pm.c | 84 +++++++++++++++++++++------------
> > > > > > 1 file changed, 55 insertions(+), 29 deletions(-)
> > > > > >
> > > > > > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> > > > > > index f8d62d1977ac..27d4d626cb34 100644
> > > > > > --- a/drivers/gpu/drm/i915/intel_pm.c
> > > > > > +++ b/drivers/gpu/drm/i915/intel_pm.c
> > > > > > @@ -3757,42 +3757,25 @@ intel_disable_sagv(struct drm_i915_private *dev_priv)
> > > > > > return 0;
> > > > > > }
> > > > > >
> > > > > > -bool intel_can_enable_sagv(struct intel_atomic_state *state)
> > > > > > +static bool icl_crtc_can_enable_sagv(const struct intel_crtc_state *crtc_state)
> > > > > > {
> > > > > > - struct drm_device *dev = state->base.dev;
> > > > > > - struct drm_i915_private *dev_priv = to_i915(dev);
> > > > > > - struct intel_crtc *crtc;
> > > > > > + struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
> > > > > > + struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
> > > > > > struct intel_plane *plane;
> > > > > > - struct intel_crtc_state *crtc_state;
> > > > > > - enum pipe pipe;
> > > > > > + const struct intel_plane_state *plane_state;
> > > > > > int level, latency;
> > > > > >
> > > > > > - if (!intel_has_sagv(dev_priv))
> > > > > > + if (crtc_state->hw.adjusted_mode.flags & DRM_MODE_FLAG_INTERLACE) {
> > > > > > + DRM_DEBUG_KMS("No SAGV for interlaced mode on pipe %c\n",
> > > > > > + pipe_name(crtc->pipe));
> > > > > > return false;
> > > > > > + }
> > > > > >
> > > > > > - /*
> > > > > > - * If there are no active CRTCs, no additional checks need be performed
> > > > > > - */
> > > > > > - if (hweight8(state->active_pipes) == 0)
> > > > > > + if (!crtc_state->hw.active)
> > > > >
> > > > > Should really be checked before anything else. Doesn't matter too much
> > > > > anymore since I made us clear the crtc state always, but still a bit
> > > > > inconsistent to look at other stuff in the state before we even know if
> > > > > the crtc is even enabled.
> > > > >
> > > > > > return true;
> > > > > >
> > > > > > - /*
> > > > > > - * SKL+ workaround: bspec recommends we disable SAGV when we have
> > > > > > - * more then one pipe enabled
> > > > > > - */
> > > > > > - if (hweight8(state->active_pipes) > 1)
> > > > > > - return false;
> > > > > > -
> > > > > > - /* Since we're now guaranteed to only have one active CRTC... */
> > > > > > - pipe = ffs(state->active_pipes) - 1;
> > > > > > - crtc = intel_get_crtc_for_pipe(dev_priv, pipe);
> > > > > > - crtc_state = to_intel_crtc_state(crtc->base.state);
> > > > > > -
> > > > > > - if (crtc_state->hw.adjusted_mode.flags & DRM_MODE_FLAG_INTERLACE)
> > > > > > - return false;
> > > > > > -
> > > > > > - for_each_intel_plane_on_crtc(dev, crtc, plane) {
> > > > > > - struct skl_plane_wm *wm =
> > > > > > + intel_atomic_crtc_state_for_each_plane_state(plane, plane_state, crtc_state) {
> > > > > > + const struct skl_plane_wm *wm =
> > > > > > &crtc_state->wm.skl.optimal.planes[plane->id];
> > > > > >
> > > > > > /* Skip this plane if it's not enabled */
> > > > > > @@ -3807,7 +3790,7 @@ bool intel_can_enable_sagv(struct intel_atomic_state *state)
> > > > > > latency = dev_priv->wm.skl_latency[level];
> > > > > >
> > > > > > if (skl_needs_memory_bw_wa(dev_priv) &&
> > > > > > - plane->base.state->fb->modifier ==
> > > > > > + plane_state->uapi.fb->modifier ==
> > > > >
> > > > > hw.fb
> > > > >
> > > > > With those this is basically good, but still need to think how to avoid
> > > > > the regression due to only checking the crtcs in the state.
> > > >
> > > > Well tbh, initially you told me that this *_crtc_can_enable_sagv function extraction
> > > > can be "trivially" done as a separate patch :)) So I followed your instruction, and
> > > > then I got a comment saying that "this is now temporary busted because we are checking
> > > > only crtcs in a state". This kind of contraversial requirements - in order not to
> > > > have it "temporary busted", we should have introduced it at the same time with SAGV mask,
> > > > at the same time you wanted it to be extracted as a separate patch.
> > >
> > > TBF this patch does quite a bit more than extract the current code.
> > >
> > > What I think would work as a series is something like:
> > > patch 1:
> > > +intel_crtc_can_enable_sagv(crtc_state)
> > > {
> > > + stuff
> > > }
> > >
> > > intel_can_enable_sagv(state)
> > > {
> > > ...
> > > crtc_state = to_intel_crtc_state(crtc->base.state);
> > >
> > > - stuff
> > > + return intel_crtc_can_eanble_sagv(crtc_state);
> > > }
> > >
> > > patch 2:
> > > +sagv_pre_plane_update(state)
> > > +{
> > > + if (!intel_can_enable_sagv(state))
> > > + intel_disable_sagv(dev_priv);
> > > +}
> > >
> > > intel_atomic_commit_tail()
> > > {
> > > ...
> > > - if (!intel_can_enable_sagv(state))
> > > - intel_disable_sagv(dev_priv);
> > > + sagv_pre_plane_update(state);
> > > ...
> > > }
> > >
> > > (+ identical changes for post_plane_update())
> > >
> > > So far everything has been pure refactoring.
> > >
> > > patch 3:
> > > Introduce the sagv mask in bw state and precompute it using
> > > intel_crtc_can_enable_sagv() (while fixing the iterator issue therein),
> > > and update the pre/post hooks to consult said mask. Not quite pure
> > > refactoring anymore but seems like a bit more straightforward change
> > > now.
> > >
> > > At this point we should have a nicely precomputed sagv mask without
> > > intentional changes to current behaviour. After which it should be
> > > easier to extend this for new platforms.
> >
> > This all makes sense, however in the end we'll have the same result as now, however this would
> > require to reshuffle the whole series...again.
> > Will try do it, the least painful way :)
>
> Also the only weird thing with this approach is that it is going to stay
> this ugly _one_ crtc way, until sagv mask and bw state is introduced:
>
> bool intel_can_enable_sagv(struct intel_atomic_state *state)
> {
> struct drm_i915_private *dev_priv = to_i915(state->base.dev);
> struct intel_crtc *crtc;
> const struct intel_crtc_state *crtc_state;
> int i;
>
> if (!intel_has_sagv(dev_priv))
> return false;
>
> /*
> * SKL+ workaround: bspec recommends we disable SAGV when we have
> * more then one pipe enabled
> */
> if (hweight8(state->active_pipes) > 1)
> return false;
>
> /* Since we're now guaranteed to only have one active CRTC... */
> pipe = ffs(state->active_pipes) - 1;
> crtc = intel_get_crtc_for_pipe(dev_priv, pipe);
> crtc_state = to_intel_crtc_state(crtc->base.state);
>
> return intel_crtc_can_enable_sagv(crtc_state);
> }
>
> because you can't iterate crtcs anyway so that patch would be
> just a name change basically.
> The I can add pre/post plane update and only once bw state->sagv_mask
> is in place - the real SAGV changes can come. So SAGV logic would be
> anyway wrong in the middle of that series.
No more wrong than it is now.
--
Ville Syrjälä
Intel
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
