Re: [PATCH v6 7/9] drm/msm/dpu: add support for virtual planes

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 





On 11/1/2024 2:27 PM, Abhinav Kumar wrote:


On 11/1/2024 1:53 PM, Dmitry Baryshkov wrote:
On Fri, Nov 01, 2024 at 01:37:03PM -0700, Abhinav Kumar wrote:


On 10/31/2024 2:03 PM, Dmitry Baryshkov wrote:
On Thu, Oct 31, 2024 at 01:06:34PM -0700, Abhinav Kumar wrote:


On 10/31/2024 8:11 AM, Dmitry Baryshkov wrote:
Hi Abhinav,

On Wed, 30 Oct 2024 at 21:26, Abhinav Kumar <quic_abhinavk@xxxxxxxxxxx> wrote:



On 10/30/2024 3:48 AM, Dmitry Baryshkov wrote:
On Tue, Oct 29, 2024 at 02:30:12PM -0700, Abhinav Kumar wrote:


On 10/24/2024 5:20 PM, Dmitry Baryshkov wrote:
Only several SSPP blocks support such features as YUV output or scaling, thus different DRM planes have different features.  Properly utilizing
all planes requires the attention of the compositor, who should
prefer simpler planes to YUV-supporting ones. Otherwise it is very easy
to end up in a situation when all featureful planes are already
allocated for simple windows, leaving no spare plane for YUV playback.

To solve this problem make all planes virtual. Each plane is registered as if it supports all possible features, but then at the runtime during the atomic_check phase the driver selects backing SSPP block for each
plane.

As the planes are attached to the CRTC and not the encoder, the SSPP blocks are also allocated per CRTC ID (all other resources are currently allocated per encoder ID). This also matches the hardware requirement, where both rectangles of a single SSPP can only be used with the LM
pair.

Note, this does not provide support for using two different SSPP blocks for a single plane or using two rectangles of an SSPP to drive two planes. Each plane still gets its own SSPP and can utilize either a solo rectangle or both multirect rectangles depending on the resolution.

Note #2: By default support for virtual planes is turned off and the driver still uses old code path with preallocated SSPP block for each plane. To enable virtual planes, pass 'msm.dpu_use_virtual_planes=1'
kernel parameter.

Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx>
---
      drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c  |  50 +++++++
      drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c   |  10 +-
      drivers/gpu/drm/msm/disp/dpu1/dpu_kms.h   |   4 +
      drivers/gpu/drm/msm/disp/dpu1/dpu_plane.c | 237 ++++++++++++++++++++++++++----
      drivers/gpu/drm/msm/disp/dpu1/dpu_plane.h |  16 ++
      drivers/gpu/drm/msm/disp/dpu1/dpu_rm.c    |  68 +++++++++
      drivers/gpu/drm/msm/disp/dpu1/dpu_rm.h    |  27 ++++
      7 files changed, 383 insertions(+), 29 deletions(-)

diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c
index 58595dcc3889..a7eea094aa14 100644
--- a/drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c
+++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c
@@ -1166,6 +1166,49 @@ static bool dpu_crtc_needs_dirtyfb(struct drm_crtc_state *cstate)
       return false;
      }
+static int dpu_crtc_reassign_planes(struct drm_crtc *crtc, struct drm_crtc_state *crtc_state)
+{
+   int total_planes = crtc->dev->mode_config.num_total_plane;
+   struct drm_atomic_state *state = crtc_state->state;
+   struct dpu_global_state *global_state;
+   struct drm_plane_state **states;
+   struct drm_plane *plane;
+   int ret;
+
+   global_state = dpu_kms_get_global_state(crtc_state->state);
+   if (IS_ERR(global_state))
+           return PTR_ERR(global_state);
+
+   dpu_rm_release_all_sspp(global_state, crtc);
+
+   if (!crtc_state->enable)
+           return 0;
+
+   states = kcalloc(total_planes, sizeof(*states), GFP_KERNEL);
+   if (!states)
+           return -ENOMEM;
+
+   drm_atomic_crtc_state_for_each_plane(plane, crtc_state) {
+           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;
+           }
+
+           states[plane_state->normalized_zpos] = plane_state;
+   }
+
+   ret = dpu_assign_plane_resources(global_state, state, crtc, states, total_planes);
+
+done:
+   kfree(states);
+   return ret;
+
+   return 0;
+}
+
      static int dpu_crtc_atomic_check(struct drm_crtc *crtc,
               struct drm_atomic_state *state)
      {
@@ -1181,6 +1224,13 @@ static int dpu_crtc_atomic_check(struct drm_crtc *crtc,
       bool needs_dirtyfb = dpu_crtc_needs_dirtyfb(crtc_state);
+   if (dpu_use_virtual_planes &&
+       (crtc_state->planes_changed || crtc_state->zpos_changed)) {
+           rc = dpu_crtc_reassign_planes(crtc, crtc_state);
+           if (rc < 0)
+                   return rc;
+   }

planes_changed is set only for format changes . Will it cover all
needs_modeset cases?

OR do we also need to set planes_changed when
drm_atomic_crtc_needs_modeset()?

Unless I am missing something, I think we have to otherwise sspp
reallocation wont happen in modeset cases.

I was depending on the planes being included in the state by the client. I don't think we really care about the modeset per se. We care about
plane size changes. And changing the size means that the plane is
included into the commit.


The global state mapping for SSPPs has to be cleared across modesets
IMO. This is no different from us calling dpu_rm_release() today in
dpu_encoder_virt_atomic_check(). I just am not sure whether
planes_changed will cover all modeset conditions.

We clear other resources, because they depend on the CRTC resolution.
Planes do not. Well, not until the quadpipe is in play.
SSPPs (currently) should be reallocated only if the _plane_'s
resolution change. If we have a modeset which involves CRTC resolution
change, but not the plane's size change, there is no need to
reallocate SSPPs.


In dpu_encoder_helper_phys_cleanup(), the SSPPs attached to all LMs are removed so clearing all the hardware. If the global state is still going to retain the older configuration not reflecting this clear, it seems incorrect
to me. Thats why I was thinking of clearing all the SSPP mapping in
disable() or in the modeset prior to the disable as technically thats being
done in HW today anyway.

During the next atomic check, the planes in the crtc's current state will get re-attached and programmed to the blend stages. So this clearing of global state is reflecting the current state of the corresponding hardware.

The global state tracks resource allocation. If we clear the resources
in the disable() path, we have no way to know which SSPP blocks were
assigned to us in the corresponding enable() call path. There is no
guarantee that there will be an atomic_check() between disable() and
enable().


So I had suggested clearing in disable() because we did not come to an
agreement to doing it in atomic_check() just a few comments earlier.

Doing it in disable() is not right. I agree with that part now as we should
not be touching the state after atomic_check() phase.

That brings me back to my original question. With the planes_changed check in atomic_check how can we guarantee that global state SSPP allocation is freed and allocated again across a disable() / enable() cycle? Can you pls confirm whether this is happening or not across a hotplug and suspend/resume
cycle?

disable() / enable() on which object? Because CRTC, if it
needs_modeset() || crtc_needs_disable() absolutely can go through a
disable / enable cycle, it doesn't require SSPP reallocation at all.


This is the part I am failing to understand. So as I wrote above, across a disable() / enable() cycle the SSPPs are cleared from the LMs and re-attached on the commit which enables() the CRTC back.

All that I am saying is that the global state SSPP mapping to the crtc_id should also reflect this clearing. Otherwise this will lead to a mismatch of states.


But maybe it's easier to just have drm_atomic_crtc_needs_modeset(). Will
that make it better for you?


Yes this is exactly what I had requested in my first response on this thread.


Summarizing our discussion from IRC:

I was more interested in the case where there is only one SSPP and lets say it was bound to CRTC0. Then CRTC0 gets disabled and CRTC1 gets enabled. Then we will have to free up the SSPP from CRTC0.

But based on our discussion, drm_atomic_helper_disable_plane() should get called which in-turn will lead to planes_chaged to be set.

With this in mind, from the code review standpoint, I cannot think of basic test-cases which can get broken, hence,

Reviewed-by: Abhinav Kumar <quic_abhinavk@xxxxxxxxxxx>

From validation standpoint, I would certainly like to test more myself but nothing immediately which i can hold back this change for.





Were you able to confirm whether the mapping gets cleared across
hotplugs or suspend/resumes? If so, it would confirm whether
planes_changed covers these aspects. Although, I think clearing should
be more explicit.

I will check that tomorrow.

Another option could be for you to call dpu_rm_release_all_sspp() in
dpu_crtc_disable(). So that across a disable and enable we have a clear
mapping table. WDYT?

Absolutely no. The RM state should only be changed when other object's
state change - in atomic_check(). After that it is mostly r/o.
enabling/disabling the resource shouldn't change resource assignment
at all.


Ack but please check above.



Overall, mainly we want to make sure SSPPs are re-assigned when:


0) plane size changes
1) format changes (RGB to YUV and vice-versa)
2) Any modesets

No

I am not able to follow why this is different from any global state
mapping of other HW blocks that we do across modesets.

DIfferent lifecycle requirements, I'd say.



3) Any disable/enable without modeset like connectors changed as SSPPs are
changing outputs there.

Absolutely no, the logic should be the same as active vs enabled for
CRTCs. Realloc resources only if the plane itself gets disabled or
enabled. In all other cases the set of SSPP blocks should stay
untouched.


I am going to re-visit this later perhaps but if we incorporate my above
suggestion of clearing the mapping in disable() I will be partially
satisfied.

No, resource mapping can not be cleaned in disable(). We do not do
that for any other resource kind.


That gets handled with the needs_modeset part today which is calling the
dpu_rm_release().

In atomic_check() path, not in the disable() path.




If we are covered for all these, let me know.

+
       if (!crtc_state->enable || !drm_atomic_crtc_effectively_active(crtc_state)) {                DRM_DEBUG_ATOMIC("crtc%d -> enable %d, active %d, skip atomic_check\n",
                               crtc->base.id, crtc_state->enable,
diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c
index 15679dd50c66..70757d876cc3 100644
--- a/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c
+++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c
@@ -51,6 +51,9 @@
      #define DPU_DEBUGFS_DIR "msm_dpu"
      #define DPU_DEBUGFS_HWMASKNAME "hw_log_mask"
+bool dpu_use_virtual_planes;
+module_param(dpu_use_virtual_planes, bool, 0);
+
      static int dpu_kms_hw_init(struct msm_kms *kms);
      static void _dpu_kms_mmu_destroy(struct dpu_kms *dpu_kms);
@@ -814,8 +817,11 @@ static int _dpu_kms_drm_obj_init(struct dpu_kms *dpu_kms)
                         type, catalog->sspp[i].features,
                         catalog->sspp[i].features & BIT(DPU_SSPP_CURSOR)); -           plane = dpu_plane_init(dev, catalog->sspp[i].id, type,
-                                  (1UL << max_crtc_count) - 1);
+           if (dpu_use_virtual_planes)
+                   plane = dpu_plane_init_virtual(dev, type, (1UL << max_crtc_count) - 1);
+           else
+                   plane = dpu_plane_init(dev, catalog->sspp[i].id, type, +                                          (1UL << max_crtc_count) - 1);
               if (IS_ERR(plane)) {
                       DPU_ERROR("dpu_plane_init failed\n");
                       ret = PTR_ERR(plane);
diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.h b/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.h
index 935ff6fd172c..479d4c172290 100644
--- a/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.h
+++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.h
@@ -54,6 +54,8 @@
      #define ktime_compare_safe(A, B) \
       ktime_compare(ktime_sub((A), (B)), ktime_set(0, 0))
+extern bool dpu_use_virtual_planes;
+
      struct dpu_kms {
       struct msm_kms base;
       struct drm_device *dev;
@@ -128,6 +130,8 @@ struct dpu_global_state {
       uint32_t dspp_to_enc_id[DSPP_MAX - DSPP_0];
       uint32_t dsc_to_enc_id[DSC_MAX - DSC_0];
       uint32_t cdm_to_enc_id;
+
+   uint32_t sspp_to_crtc_id[SSPP_MAX - SSPP_NONE];
      };

This is the part which now looks odd and can be managed with rebase I guess.

Are you planning to pull in the move resource allocation to crtc_id changes
first before this part? IOW, rebase this change on top of that?

No. I do not. If you remember, several revisions ago the enc_id ->
crtc_id was a part of the series, but we both agreed to drop it since it was not required for virtual planes. As such, I plan to land this one first (yes, having some of the resources tracked basing on enc_id and
SSPP is tracked basing on crtc_id).


Yes, I am not asking whether you will be absorbing those changes into
this series. Even I would not suggest doing that.

I was asking whether you will merge the crtc_id based tracking first and
then apply this on top of that and not the other way around.

Because with this specific line I am certain it will conflict as both
the series touch struct dpu_global_state.

They touch different parts of it. So I'd prefer to land this one first
and then land using crtc_id for mapping.


I am okay to fixup any other issues which arise later on because we have the modparam protection anyway but I think validating suspend/resume and hotplug to ensure no black screens is required. If those two cases work fine on your
end, we can proceed.

I have been validating these changes with hotplug events, yes. I wasn't
checking the suspend/resume, but that's broken anyway, until we land
https://patchwork.freedesktop.org/patch/606931/?series=135908&rev=2


Can you pls confirm once whether the global state mapping gets freed across crtc disable/enable cycle with the planes_changed check? I think it has to.

I think you are asking the question from the wrong side. What kind of
commit leads to that CRTC disable/enable cycle?


A commit which requires modeset?


Other items are closed so snipping out below.




[Index of Archives]     [Linux DRI Users]     [Linux Intel Graphics]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]
  Powered by Linux