On Tue, May 24, 2016 at 05:01:02PM +0300, Laurent Pinchart wrote:
Hi Ville,
On Thursday 12 May 2016 23:54:09 Ville Syrjälä wrote:
On Thu, May 12, 2016 at 11:15:42PM +0300, Laurent Pinchart wrote:
On Thursday 12 May 2016 22:46:05 Ville Syrjälä wrote:
On Thu, May 12, 2016 at 10:20:12PM +0300, Laurent Pinchart wrote:
On Thursday 12 May 2016 14:51:55 Ville Syrjälä wrote:
On Thu, May 12, 2016 at 12:28:19PM +0200, Benjamin Gaignard wrote: > From: Marek Szyprowski m.szyprowski@samsung.com > > This patch adds support for generic plane's zpos property property > with well-defined semantics: > - added zpos properties to plane and plane state structures > - added helpers for normalizing zpos properties of given set of > planes > - well defined semantics: planes are sorted by zpos values and then > plane id value if zpos equals > > Normalized zpos values are calculated automatically when generic > muttable zpos property has been initialized. Drivers can simply use > plane_state->normalized_zpos in their atomic_check and/or > plane_update callbacks without any additional calls to DRM core. > > Signed-off-by: Marek Szyprowski m.szyprowski@samsung.com > > Compare to Marek's original patch zpos property is now specific to > each plane and no more to the core. > Normalize function take care of the range of per plane defined range > before set normalized_zpos. > > Signed-off-by: Benjamin Gaignard benjamin.gaignard@linaro.org > > Cc: Inki Dae inki.dae@samsung.com > Cc: Daniel Vetter daniel@ffwll.ch > Cc: Ville Syrjala ville.syrjala@linux.intel.com > Cc: Joonyoung Shim jy0922.shim@samsung.com > Cc: Seung-Woo Kim sw0312.kim@samsung.com > Cc: Andrzej Hajda a.hajda@samsung.com > Cc: Krzysztof Kozlowski k.kozlowski@samsung.com > Cc: Bartlomiej Zolnierkiewicz b.zolnierkie@samsung.com > Cc: Tobias Jakobi tjakobi@math.uni-bielefeld.de > Cc: Gustavo Padovan gustavo@padovan.org > Cc: vincent.abriou@st.com > Cc: fabien.dessenne@st.com > Cc: Laurent Pinchart laurent.pinchart@ideasonboard.com > --- > > Documentation/DocBook/gpu.tmpl | 10 ++ > drivers/gpu/drm/Makefile | 2 +- > drivers/gpu/drm/drm_atomic.c | 4 + > drivers/gpu/drm/drm_atomic_helper.c | 6 + > drivers/gpu/drm/drm_blend.c | 229 +++++++++++++++++++++++++ > drivers/gpu/drm/drm_crtc_internal.h | 3 + > include/drm/drm_crtc.h | 28 +++++ > 7 files changed, 281 insertions(+), 1 deletion(-) > create mode 100644 drivers/gpu/drm/drm_blend.c
[snip]
> diff --git a/drivers/gpu/drm/drm_blend.c > b/drivers/gpu/drm/drm_blend.c > new file mode 100644 > index 0000000..eef59bb > --- /dev/null > +++ b/drivers/gpu/drm/drm_blend.c > @@ -0,0 +1,229 @@
[snip]
> +static int drm_atomic_state_zpos_cmp(const void *a, const void *b) > +{ > + const struct drm_plane_state *sa = *(struct drm_plane_state **)a; > + const struct drm_plane_state *sb = *(struct drm_plane_state **)b; > + > + if (sa->zpos != sb->zpos) > + return sa->zpos - sb->zpos; > + else > + return sa->plane->base.id - sb->plane->base.id; > +}
Still not really convinced this normalization is a good idea when we have atomic. It made sense before atomic since otherwise you might not be able to swap the zpos of two planes without shutting one of them down, but it makes interpreting the zpos prop values more confusing. And with atomic we don't actually need it AFAICS. Or do we have an actual use case for this behaviour?
Speaking selfishly about my use case (rcar-du-drm), I need to sort planes in display order from 0 to N-1 as that's how the hardware is programmed. Normalization gives me an array of sorted planes, which is exactly what I need. I expect other drivers to have similar requirements, so normalization in the core will allow code sharing.
It just eliminates duplicate zpos values. It won't re-sort your plane states magically or anything like that. Nor does it guarantee that the zpos values for the actually visible planes are even contiguous.
To make normalization useful in my case I need contiguous values starting at 0.
Contigious values for all the planes on the whole device or on one crtc? And what about the visible vs. not plane issue? Is it OK to have an invisible plane somewhere in the middle of the zpos sequence?
In my case it would be per-CRTC, and invisible planes in the middle wouldn't be an issue.
To make normalization useful it needs to compute values that are usable by the majority of drivers. The question then becomes what should those values be ? Generally speaking they should be easy to convert to driver-specific values. That's why I believe we need feedback from driver developers on this topic.
Now, is your concern that allowing two planes to have the same zpos isn't needed with the atomic API ? I agree with that, but it would also make it impossible to swap two planes using the legacy API implemented on top of atomic drivers. I'll let others decide whether this is considered as a concern.
Any opinion on this ?
Not really. I already asked if we had a good use case for it. Dunno if anyone has one.
Non-atomic userspace needing to swap planes is such a use case. The question is whether we want to support it.
I was more asking if anyone is actually doing that, and do they need it to keep working.
For atomic, the reason I can see to support the same feature is to make it easier for userspace to modify the zorder of a plane without requiring updates to all other planes. I don't have a strong opinion on that, but we need to decide whether we want to support it or not.
Not sure it really helps in any measurable way. It might save a few properties from getting pushed to the kernel, but that's all. Eg. if userspace would maintain some kind of stack of planes in the right zorder, then pushing a new plane into the middle of the stack would required that it figures out the right zorder for that position. That would be something like z=clamp(below_z + 1, below_z, above_z). After that it would have to figure out if the magic conflict resolution will result in the right zorder should below_z==z or above_z==z, and if not it would have to change below_z and/or above_z too. That process could ripple all the way through the entire stack. Seems to me that it would be far easier to just readjust the z order for all the planes above from the get go.
While at it, Benjamin told me that you've requested configurable min/max values per plane, what are your use cases for that ? On the same topic, please see below.
> + > +/** > + * drm_atomic_helper_crtc_normalize_zpos - calculate normalized > zpos values > + * @crtc: crtc with planes, which have to be considered for > normalization > + * @crtc_state: new atomic state to apply > + * > + * This function checks new states of all planes assigned to given > crtc and > + * calculates normalized zpos value for them. Planes are compared > first by their > + * zpos values, then by plane id (if zpos equals). Plane with > lowest zpos value > + * is at the bottom. The plane_state->normalized_zpos is then > filled with unique > + * values from 0 to number of active planes in crtc minus one. > + * > + * RETURNS > + * Zero for success or -errno > + */ > +int drm_atomic_helper_crtc_normalize_zpos(struct drm_crtc *crtc, > + struct drm_crtc_state *crtc_state) > +{ > + struct drm_atomic_state *state = crtc_state->state; > + struct drm_device *dev = crtc->dev; > + int total_planes = dev->mode_config.num_total_plane; > + struct drm_plane_state **states; > + struct drm_plane *plane; > + int i, zpos, n = 0; > + int ret = 0; > + > + DRM_DEBUG_ATOMIC("[CRTC:%d:%s] calculating normalized zpos > values\n", > + crtc->base.id, crtc->name); > + > + states = kmalloc(total_planes * sizeof(*states), GFP_KERNEL);
kmalloc_array(..., GFP_TEMPORARY);
> + if (!states) > + return -ENOMEM; > + > + /* > + * Normalization process might create new states for planes which > + * normalized_zpos has to be recalculated. > + */ > + drm_for_each_plane_mask(plane, dev, crtc_state->plane_mask) { > + struct drm_plane_state *plane_state = > + drm_atomic_get_plane_state(state, plane); > + if (IS_ERR(plane_state)) { > + ret = PTR_ERR(plane_state); > + goto done; > + }
Hmm. So if we change the zpos for any plane, we get to re-check every plane. Not the most efficient thing perhaps, since on some hardware you might only have a subset of planes that can even change their zpos. OTOH I suppose it does make life a bit simpler when you don't have to think which planes might be affected, so I guess it's fine.
> + states[n++] = plane_state; > + DRM_DEBUG_ATOMIC("[PLANE:%d:%s] processing zpos value %d\n", > + plane->base.id, plane->name, > + plane_state->zpos); > + } > + > + sort(states, n, sizeof(*states), drm_atomic_state_zpos_cmp, NULL); > + > + for (i = 0, zpos = 0; i < n; i++, zpos++) { > + plane = states[i]->plane; > + > + zpos = max_t(int, zpos, states[i]->zpos);
The purpose of this code is to support the per-plane min/max zpos values. As Benjamin explained, for example if the zpos range is [0,2] for plane X and [4,6] for plane Y, we need to be sure the min zpos value of plane Y is 4. To this I replied that's exactly what I'd rather avoid. I think the core code should take care of sorting planes, taking driver- specific constraints into account when possible (too exotic constraints can always be implemented directly by drivers), and then leave to the drivers the responsibility of taking the sorted planes array and computing whatever hardware-specific configuration is needed. Sorting planes from 0 to N-1 seems to be like a better API than having semi- random values for the normalized zpos. Do we have use cases that wouldn't be covered by a simpler implementation ?
The point of exposing ranges is to give userspace at least some chance of getting the order right. Otherwise it'll just have to randomly try different orders until something works. Most hardware is quite constrained in this.
Could you educate me by giving a few real world examples ?
Well, on Intel we have the simple case on many platforms (bottom->top): primary->sprite->cursor Most PC hardware from some years ago was similar. More modern stuff from the other vendors might be more capable, not sure.
We also have another class of Intel hardware which has primary, two sprites, and a cursor. The cursor is always on top, the other three planes can be in any order amongst themselves.
The oldest Intel hardware we support is even more flexible. Those had sort of 1 proper primary plane, 2 sprite planes, 1 video overlay, and 2 cursor planes. Normally they're spread among two crtcs, but they can all be moved to a single crtc and then ordering constraints between them are fairly convoluted.
Thank you for the information.
As far as I can remember OMAP (at least 3 and maybe 4) had a totally fixed order. GFX->VID1->VID2. Perhaps they changed this later, dunno.
OMAP3 has a fixed Z-order, while OMAP4 seems to be configurable according to the omapdrm driver.
One extra complication in this is destination color keying. What that tends to do effectively is push some sprite/overlay plane below the primary and punch a hole into the primary, but otherwise the sprite/overlay sits above the primary. Not sure if we should make this sort of thing explicit in the zpos API or if we should just declare color keying somehow magic and potentially ignore some of the zpos properties.
Color-keying is a quite common feature, I think we should standardize properties to support it, which would then require defining the behaviour, including how it interacts with Z-order.
Some don't support changing the zpos at all, in which case we would set min==max.
But all this doesn't require adjusting the normalized zpos to take the original zpos constraints into account, does it ? Why can normalized_zpos
s/can/can't/
be constrained to the range 0 to N-1, giving drivers the sorted planes and then letting them convert that to device-specific values ?
Oh, I didn't even notice that it was mixing up the normalized and non-normalized coordinates. Yes, that does seem wrong to me. The fact that the ioctl would have rejected the out of range non-normalized coordinates should already guarantee that the resulting normalized coordinates end up in roughly the right order. That is, at least the order between different "zpos groups" of planes would be already respected. Eg. in the Intel "3 freely ordered planes + cursor" case there would be no way to push the cursor below any other plane.
I agree with you, and I still think that computing the zpos_normalized values in the 0 to N-1 range (N being the number of visible planes for the CRTC) would be the best option.
You can't actually know what's visible when this gets called.