[Intel-gfx] [PATCH v2 07/20] drm/i915: Rework primary plane stuff slightly.
Daniel Vetter
daniel at ffwll.ch
Wed Jul 8 02:27:12 PDT 2015
On Tue, Jul 07, 2015 at 04:02:32PM +0200, Maarten Lankhorst wrote:
> Op 07-07-15 om 13:16 schreef Daniel Vetter:
> > On Tue, Jul 07, 2015 at 09:08:18AM +0200, Maarten Lankhorst wrote:
> >> Make sure the primary plane is set up correctly. This is done by
> >> setting plane_state->src and plane_state->crtc.
> >>
> >> All non-primary planes get disabled.
> > Too terse commit message, fails to mention that this is about hw
> > readout completely. Also should mention that this removes the
> > initial_planes hack.
> >
> >> Signed-off-by: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
> >> ---
> >> drivers/gpu/drm/i915/intel_atomic.c | 7 --
> >> drivers/gpu/drm/i915/intel_display.c | 167 +++++++++++++----------------------
> >> drivers/gpu/drm/i915/intel_drv.h | 2 -
> >> 3 files changed, 60 insertions(+), 116 deletions(-)
> >>
> >> diff --git a/drivers/gpu/drm/i915/intel_atomic.c b/drivers/gpu/drm/i915/intel_atomic.c
> >> index 429677a111d5..b593612a917d 100644
> >> --- a/drivers/gpu/drm/i915/intel_atomic.c
> >> +++ b/drivers/gpu/drm/i915/intel_atomic.c
> >> @@ -98,13 +98,6 @@ int intel_atomic_check(struct drm_device *dev,
> >> return -EINVAL;
> >> }
> >>
> >> - if (crtc_state &&
> >> - crtc_state->quirks & PIPE_CONFIG_QUIRK_INITIAL_PLANES) {
> >> - ret = drm_atomic_add_affected_planes(state, &nuclear_crtc->base);
> >> - if (ret)
> >> - return ret;
> >> - }
> >> -
> >> ret = drm_atomic_helper_check_planes(dev, state);
> >> if (ret)
> >> return ret;
> >> diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
> >> index eb7c2e2819b7..fa1102392eca 100644
> >> --- a/drivers/gpu/drm/i915/intel_display.c
> >> +++ b/drivers/gpu/drm/i915/intel_display.c
> >> @@ -109,6 +109,7 @@ static void skl_init_scalers(struct drm_device *dev, struct intel_crtc *intel_cr
> >> struct intel_crtc_state *crtc_state);
> >> static int i9xx_get_refclk(const struct intel_crtc_state *crtc_state,
> >> int num_connectors);
> >> +static void intel_pre_disable_primary(struct drm_crtc *crtc);
> >>
> >> static struct intel_encoder *intel_find_encoder(struct intel_connector *connector, int pipe)
> >> {
> >> @@ -2582,11 +2583,12 @@ intel_find_initial_plane_obj(struct intel_crtc *intel_crtc,
> >> {
> >> struct drm_device *dev = intel_crtc->base.dev;
> >> struct drm_i915_private *dev_priv = dev->dev_private;
> >> - struct drm_crtc *c;
> >> - struct intel_crtc *i;
> >> struct drm_i915_gem_object *obj;
> >> - struct drm_plane *primary = intel_crtc->base.primary;
> >> struct drm_framebuffer *fb;
> >> + struct drm_plane *primary = intel_crtc->base.primary;
> >> + struct intel_plane *p;
> >> + struct intel_plane_state *plane_state =
> >> + to_intel_plane_state(primary->state);
> >>
> >> if (!plane_config->fb)
> >> return;
> >> @@ -2602,16 +2604,11 @@ intel_find_initial_plane_obj(struct intel_crtc *intel_crtc,
> >> * Failed to alloc the obj, check to see if we should share
> >> * an fb with another CRTC instead
> >> */
> >> - for_each_crtc(dev, c) {
> >> - i = to_intel_crtc(c);
> >> -
> >> - if (c == &intel_crtc->base)
> >> - continue;
> >> -
> >> - if (!i->active)
> >> + for_each_intel_plane(dev, p) {
> >> + if (p->base.type != DRM_PLANE_TYPE_PRIMARY)
> >> continue;
> >>
> >> - fb = c->primary->fb;
> >> + fb = p->base.state->fb;
> > This seems to break the sharing logic completely: We want to check primary
> > planes of all other crtcs to see whether we could merged the fb together.
> > We don't even read out plane state for non-primary planes, so the below fb
> > check will never be non-NULL.
> I thought this was about multiple planes sharing the same fb with same offset.
> And as such checking for the crtc is unnecessary, for the current crtc it will be be NULL here.
>
> This only reads out the current fb, not different ones.
>
> And sharing the same fb with other crtc's is done in intel_fbdev_init_bios.
This is about sharing the same fb but across different crtcs - bios never
enables more than the primary plane anyway. And you can't rely upon
fbdev_init_bios since that's not run at all when I915_FBDEV=n.
So yes with current code this loop here reconstruct the shared between
primary planes on different crtcs (if the stolen allocator tells us that
our range is occupied already). fbdev_init_bios just tries to create a
config matching the one the bios has set up (and then pick a suitable fb
for fbcon from the ones already allocated).
Maybe we should extract this as try_to_find_shared_fb or similar to make
the code self-explanatory?
>
> >> if (!fb)
> >> continue;
> >>
> >> @@ -2622,18 +2619,28 @@ intel_find_initial_plane_obj(struct intel_crtc *intel_crtc,
> >> }
> >> }
> >>
> >> + intel_pre_disable_primary(&intel_crtc->base);
> >> + to_intel_plane(primary)->disable_plane(primary, &intel_crtc->base);
> >> +
> >> return;
> >>
> >> valid_fb:
> >> + drm_framebuffer_reference(fb);
> >> + primary->fb = plane_state->base.fb = fb;
> >> + plane_state->base.crtc = primary->crtc = &intel_crtc->base;
> >> +
> >> + plane_state->base.src_x = plane_state->base.src_y = 0;
> >> + plane_state->base.src_w = fb->width << 16;
> >> + plane_state->base.src_h = fb->height << 16;
> >> +
> >> + plane_state->base.crtc_x = plane_state->base.src_y = 0;
> >> + plane_state->base.crtc_w = fb->width;
> >> + plane_state->base.crtc_h = fb->height;
> >> +
> >> + plane_state->visible = true;
> >> obj = intel_fb_obj(fb);
> >> if (obj->tiling_mode != I915_TILING_NONE)
> >> dev_priv->preserve_bios_swizzle = true;
> >> -
> >> - primary->fb = fb;
> >> - primary->crtc = primary->state->crtc = &intel_crtc->base;
> >> - update_state_fb(primary);
> > Do we still have other users of update_state_fb left?
> Just the page flip handler.
> >> - intel_crtc->base.state->plane_mask |= (1 << drm_plane_index(primary));
> >> - obj->frontbuffer_bits |= to_intel_plane(primary)->frontbuffer_bit;
> >> }
> >>
> >> static void i9xx_update_primary_plane(struct drm_crtc *crtc,
> >> @@ -11761,34 +11768,6 @@ static bool check_encoder_cloning(struct drm_atomic_state *state,
> >> return true;
> >> }
> >>
> >> -static void intel_crtc_check_initial_planes(struct drm_crtc *crtc,
> >> - struct drm_crtc_state *crtc_state)
> >> -{
> >> - struct intel_crtc_state *pipe_config =
> >> - to_intel_crtc_state(crtc_state);
> >> - struct drm_plane *p;
> >> - unsigned visible_mask = 0;
> >> -
> >> - drm_for_each_plane_mask(p, crtc->dev, crtc_state->plane_mask) {
> >> - struct drm_plane_state *plane_state =
> >> - drm_atomic_get_existing_plane_state(crtc_state->state, p);
> >> -
> >> - if (WARN_ON(!plane_state))
> >> - continue;
> >> -
> >> - if (!plane_state->fb)
> >> - crtc_state->plane_mask &=
> >> - ~(1 << drm_plane_index(p));
> >> - else if (to_intel_plane_state(plane_state)->visible)
> >> - visible_mask |= 1 << drm_plane_index(p);
> >> - }
> >> -
> >> - if (!visible_mask)
> >> - return;
> >> -
> >> - pipe_config->quirks &= ~PIPE_CONFIG_QUIRK_INITIAL_PLANES;
> >> -}
> >> -
> >> static int intel_crtc_atomic_check(struct drm_crtc *crtc,
> >> struct drm_crtc_state *crtc_state)
> >> {
> >> @@ -11810,10 +11789,6 @@ static int intel_crtc_atomic_check(struct drm_crtc *crtc,
> >> "[CRTC:%i] mismatch between state->active(%i) and crtc->active(%i)\n",
> >> idx, crtc->state->active, intel_crtc->active);
> >>
> >> - /* plane mask is fixed up after all initial planes are calculated */
> >> - if (pipe_config->quirks & PIPE_CONFIG_QUIRK_INITIAL_PLANES)
> >> - intel_crtc_check_initial_planes(crtc, crtc_state);
> >> -
> >> if (mode_changed && !crtc_state->active)
> >> intel_crtc->atomic.update_wm_post = true;
> >>
> >> @@ -13155,19 +13130,6 @@ intel_modeset_compute_config(struct drm_atomic_state *state)
> >> continue;
> >> }
> >>
> >> - if (pipe_config->quirks & PIPE_CONFIG_QUIRK_INITIAL_PLANES) {
> >> - ret = drm_atomic_add_affected_planes(state, crtc);
> >> - if (ret)
> >> - return ret;
> >> -
> >> - /*
> >> - * We ought to handle i915.fastboot here.
> >> - * If no modeset is required and the primary plane has
> >> - * a fb, update the members of crtc_state as needed,
> >> - * and run the necessary updates during vblank evasion.
> >> - */
> >> - }
> >> -
> >> if (!needs_modeset(crtc_state)) {
> >> if (!(pipe_config->quirks & PIPE_CONFIG_QUIRK_INHERITED_MODE))
> >> continue;
> >> @@ -15149,25 +15111,30 @@ void intel_modeset_init(struct drm_device *dev)
> >> drm_modeset_unlock_all(dev);
> >>
> >> for_each_intel_crtc(dev, crtc) {
> >> - if (!crtc->active)
> >> + struct intel_initial_plane_config plane_config;
> >> + struct drm_plane *plane;
> >> +
> >> + if (!crtc->base.state->active)
> >> continue;
> >>
> >> + /* disable all non-primary planes */
> >> + drm_for_each_plane_mask(plane, dev,
> >> + crtc->base.state->plane_mask)
> >> + if (plane->type != DRM_PLANE_TYPE_PRIMARY)
> >> + to_intel_plane(plane)->disable_plane(plane, &crtc->base);
> >> +
> >> + crtc->base.state->plane_mask &=
> >> + 1 << drm_plane_index(crtc->base.primary);
> >> +
> >> + plane_config.fb = NULL;
> >> + dev_priv->display.get_initial_plane_config(crtc,
> >> + &plane_config);
> >> +
> >> /*
> >> - * Note that reserving the BIOS fb up front prevents us
> >> - * from stuffing other stolen allocations like the ring
> >> - * on top. This prevents some ugliness at boot time, and
> >> - * can even allow for smooth boot transitions if the BIOS
> >> - * fb is large enough for the active pipe configuration.
> >> + * If the fb is shared between multiple heads, we'll
> >> + * just get the first one.
> >> */
> >> - if (dev_priv->display.get_initial_plane_config) {
> >> - dev_priv->display.get_initial_plane_config(crtc,
> >> - &crtc->plane_config);
> >> - /*
> >> - * If the fb is shared between multiple heads, we'll
> >> - * just get the first one.
> >> - */
> >> - intel_find_initial_plane_obj(crtc, &crtc->plane_config);
> >> - }
> >> + intel_find_initial_plane_obj(crtc, &plane_config);
> >> }
> >> }
> > Ok I looked at this and the readout_plane_state function and I think we
> > have a bit a confusion about responsibilities here. The big thing is that
> > only driver load cares about reconstructing plane state accurately, for
> > resume and lid notifier we just want to make sure that we update the
> > planes. And that could be achieved by unconditionally setting
> > crtc_state->planes_changed. We already have all the plane states when
> > restoring state anyway.
> On resume I force a modeset for this reason, and set the plane_mask.
> This makes sure any plane gets disabled.
>
> The modeset will also add all planes, which sets planes_changed if needed.
Yeah that's what I had in mind for resume: We need to grab all plane
states anyway (to be able to restore the old config), so at most we need
to set a planes_changed to make sure the update happens.
> A commit on its own doesn't do this, when a plane doesn't have a fb
> it won't disable it on its own because the plane's assumed to be disabled
> when old_plane_state->fb == NULL and new_plane_state->fb == NULL.
Not a problem since too many planes doesn't seem to happen in reality. At
least we didn't have code to force-disable planes on resume/lid_notify
before, which means we don't suddenly need it now. sanitize_* should only
fix up stuff that's broken for i915 on real-world machines, not what all
could be possible.
> I don't know if I can call disable_plane in this loop either, because that will
> update the watermarks on skylake with the call to intel_update_sprite_watermarks
> calling skl_update_sprite_wm calling skl_update_wm.
Hm that sounds like a bug still. Maybe just forget about disabling
non-primary planes for now on driver load? We don't seem to have any need
currently either. And for primary planes maybe we can hard-code something
which just clears the PLANE_ENABLE bit and does nothing else? Kind of a
super-low-level plane force disable. Similar to how we have a special
function to force-disable the vga plane.
> We don't even have any watermarks calculated yet, so that will break.
> > That means we should move readout_plane_state into the above loop. Which
> > then gets a bit too big, so better extract this into a
> > intel_reconstruct_plane_state or similar. This function should then do all
> > the plane state reconstruction.
> >
> > That also means we don't have to play tricks with plane_mask like you do.
> > We simply reconstruct the primary plane (if possible) and force-disable
> > all the others. Since this only happens at driver load there's no need to
> > clear out any state for sprite/cursors since it's already fully cleared.
> >
> > I think this way we should be able to have everything in one place, and
> > that should allow us to simplify things a lot.
> I do this trick for atomic resume, we can't allocate the original fb in that case
> but I still want to sanitize everything. This either happens because
> a new primary fb gets committed or the primary fb gets disabled.
Well on resume we don't care at all about the original fb for two reasons:
- Anything remotely modern resumes with everything off.
- We only recover the initial plane from the bios to ensure boot-up is
completely flicker-free. If we don't reserve that we'll allocate rings
and other gpu objects in there which isn't pretty.
I don't think we need to sanitize planes either as long as we force a full
plane update on resume (by forcing planes_changed or similar).
And for the lid_restore case the problem hasn't been that the bios changed
the plane state really, but that it just went ahead and disabled
everything on its own. That will result in a crtc_enable which will
restore everything correctly.
Imo we don't need to have a perfect sanitize for everything, this code
really just fixes up what's actually been broken in real-world machines
out there. That kind of real-world testing is also why I want to change as
little as possible in the sanitize_* functions.
> >> @@ -15404,47 +15371,30 @@ static void readout_plane_state(struct intel_crtc *crtc,
> >> struct intel_plane *p;
> >> struct drm_plane_state *drm_plane_state;
> >> bool active = crtc_state->base.active;
> >> + bool wrong_plane;
> >>
> >> - if (active) {
> >> - crtc_state->quirks |= PIPE_CONFIG_QUIRK_INITIAL_PLANES;
> >> -
> >> - /* apply to previous sw state too */
> >> - to_intel_crtc_state(crtc->base.state)->quirks |=
> >> - PIPE_CONFIG_QUIRK_INITIAL_PLANES;
> >> - }
> >> + wrong_plane = active && !intel_check_plane_mapping(crtc);
> >> + if (wrong_plane)
> >> + crtc->pipe = !crtc->pipe;
> >>
> >> for_each_intel_plane(crtc->base.dev, p) {
> >> - bool visible = active;
> >> -
> >> if (crtc->pipe != p->pipe)
> >> continue;
> >>
> >> drm_plane_state = p->base.state;
> >>
> >> - /* Plane scaler state is not touched here. The first atomic
> >> - * commit will restore all plane scalers to its old state.
> >> - */
> >> -
> >> - if (active && p->base.type == DRM_PLANE_TYPE_PRIMARY) {
> >> - visible = primary_get_hw_state(crtc);
> >> - to_intel_plane_state(drm_plane_state)->visible = visible;
> >> - } else {
> >> - /*
> >> - * unknown state, assume it's off to force a transition
> >> - * to on when calculating state changes.
> >> - */
> >> + if (p->base.type == DRM_PLANE_TYPE_PRIMARY)
> >> + to_intel_plane_state(drm_plane_state)->visible =
> >> + primary_get_hw_state(crtc);
> >> + else
> >> to_intel_plane_state(drm_plane_state)->visible = false;
> >> - }
> >>
> >> - if (visible) {
> >> - crtc_state->base.plane_mask |=
> >> - 1 << drm_plane_index(&p->base);
> >> - } else if (crtc_state->base.state) {
> >> - /* Make this unconditional for atomic hw readout. */
> >> - crtc_state->base.plane_mask &=
> >> - ~(1 << drm_plane_index(&p->base));
> >> - }
> >> + crtc_state->base.plane_mask |=
> >> + 1 << drm_plane_index(&p->base);
> >> }
> >> +
> >> + if (wrong_plane)
> >> + crtc->pipe = !crtc->pipe;
> >> }
> >>
> >> static void intel_modeset_readout_hw_state(struct drm_device *dev)
> >> @@ -15664,6 +15614,9 @@ void intel_modeset_gem_init(struct drm_device *dev)
> >> update_state_fb(c->primary);
> >> c->state->plane_mask &= ~(1 << drm_plane_index(c->primary));
> >> }
> >> + else
> >> + obj->frontbuffer_bits |=
> >> + to_intel_plane(c->primary)->frontbuffer_bit;
> > This is part of plane state reconstruction, no need to move it. The only
> > reason we have to pin late is that gem isn't initialized yet fully when we
> > want to reconstruct the modeset state. And that pinning shouldn't ever
> > fail, it just increments pin_count and writes the ptes (again).
> Why bother handling the return value at all then?
To print the warning. I think even the cleanup is pointless since at worst
we'll underrun the pin count, and we have protection for that already
(again with a loud warning). If you want you can remove that code.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
More information about the Intel-gfx
mailing list