drm/amd/display: Use vblank control events for PSR enable/disable
authorNicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Tue, 3 Aug 2021 20:32:31 +0000 (16:32 -0400)
committerAlex Deucher <alexander.deucher@amd.com>
Mon, 16 Aug 2021 19:35:56 +0000 (15:35 -0400)
[Why]
PSR can disable the HUBP along with the OTG when PSR is active.

We'll hit a pageflip timeout when the OTG is disable because we're no
longer updating the CRTC vblank counter and the pflip high IRQ will
not fire on the flip.

In order to flip the page flip timeout occur we should modify the
enter/exit conditions to match DRM requirements.

[How]
Use our deferred handlers for DRM vblank control to notify DMCU(B)
when it can enable or disable PSR based on whether vblank is disabled or
enabled respectively.

We'll need to pass along the stream with the notification now because
we want to access the CRTC state while the CRTC is locked to get the
stream state prior to the commit.

Retain a reference to the stream so it remains safe to continue to
access and release that reference once we're done with it.

Enable/disable logic follows what we were previously doing in
update_planes.

The workqueue has to be flushed before programming streams or planes
to ensure that we exit out of idle optimizations and PSR before
these events occur if necessary.

To keep the skip count logic the same to avoid FBCON PSR enablement
requires copying the allow condition onto the DM IRQ parameters - a
field that we can actually access from the worker.

Reviewed-by: Roman Li <Roman.Li@amd.com>
Acked-by: Wayne Lin <wayne.lin@amd.com>
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Tested-by: Daniel Wheeler <daniel.wheeler@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.h
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_irq_params.h

index f88b6c5b83cd2d53651758e14f99f5365759bcb8..cebd663b670888d849ad509a3ca27d60d596a506 100644 (file)
@@ -1061,7 +1061,22 @@ static void vblank_control_worker(struct work_struct *work)
 
        DRM_DEBUG_KMS("Allow idle optimizations (MALL): %d\n", dm->active_vblank_irq_count == 0);
 
+       /* Control PSR based on vblank requirements from OS */
+       if (vblank_work->stream && vblank_work->stream->link) {
+               if (vblank_work->enable) {
+                       if (vblank_work->stream->link->psr_settings.psr_allow_active)
+                               amdgpu_dm_psr_disable(vblank_work->stream);
+               } else if (vblank_work->stream->link->psr_settings.psr_feature_enabled &&
+                          !vblank_work->stream->link->psr_settings.psr_allow_active &&
+                          vblank_work->acrtc->dm_irq_params.allow_psr_entry) {
+                       amdgpu_dm_psr_enable(vblank_work->stream);
+               }
+       }
+
        mutex_unlock(&dm->dc_lock);
+
+       dc_stream_release(vblank_work->stream);
+
        kfree(vblank_work);
 }
 
@@ -6018,6 +6033,11 @@ static inline int dm_set_vblank(struct drm_crtc *crtc, bool enable)
        work->acrtc = acrtc;
        work->enable = enable;
 
+       if (acrtc_state->stream) {
+               dc_stream_retain(acrtc_state->stream);
+               work->stream = acrtc_state->stream;
+       }
+
        queue_work(dm->vblank_control_workqueue, &work->work);
 #endif
 
@@ -8623,6 +8643,12 @@ static void amdgpu_dm_commit_planes(struct drm_atomic_state *state,
        /* Update the planes if changed or disable if we don't have any. */
        if ((planes_count || acrtc_state->active_planes == 0) &&
                acrtc_state->stream) {
+               /*
+                * If PSR or idle optimizations are enabled then flush out
+                * any pending work before hardware programming.
+                */
+               flush_workqueue(dm->vblank_control_workqueue);
+
                bundle->stream_update.stream = acrtc_state->stream;
                if (new_pcrtc_state->mode_changed) {
                        bundle->stream_update.src = acrtc_state->stream->src;
@@ -8691,16 +8717,20 @@ static void amdgpu_dm_commit_planes(struct drm_atomic_state *state,
                                acrtc_state->stream->link->psr_settings.psr_version != DC_PSR_VERSION_UNSUPPORTED &&
                                !acrtc_state->stream->link->psr_settings.psr_feature_enabled)
                        amdgpu_dm_link_setup_psr(acrtc_state->stream);
-               else if ((acrtc_state->update_type == UPDATE_TYPE_FAST) &&
-                               acrtc_state->stream->link->psr_settings.psr_feature_enabled &&
-                               !acrtc_state->stream->link->psr_settings.psr_allow_active) {
-                       struct amdgpu_dm_connector *aconn = (struct amdgpu_dm_connector *)
-                                       acrtc_state->stream->dm_stream_context;
+
+               /* Decrement skip count when PSR is enabled and we're doing fast updates. */
+               if (acrtc_state->update_type == UPDATE_TYPE_FAST &&
+                   acrtc_state->stream->link->psr_settings.psr_feature_enabled) {
+                       struct amdgpu_dm_connector *aconn =
+                               (struct amdgpu_dm_connector *)acrtc_state->stream->dm_stream_context;
 
                        if (aconn->psr_skip_count > 0)
                                aconn->psr_skip_count--;
-                       else
-                               amdgpu_dm_psr_enable(acrtc_state->stream);
+
+                       /* Allow PSR when skip count is 0. */
+                       acrtc_attach->dm_irq_params.allow_psr_entry = !aconn->psr_skip_count;
+               } else {
+                       acrtc_attach->dm_irq_params.allow_psr_entry = false;
                }
 
                mutex_unlock(&dm->dc_lock);
@@ -8949,8 +8979,10 @@ static void amdgpu_dm_atomic_commit_tail(struct drm_atomic_state *state)
 
        if (dc_state) {
                /* if there mode set or reset, disable eDP PSR */
-               if (mode_set_reset_required)
+               if (mode_set_reset_required) {
+                       flush_workqueue(dm->vblank_control_workqueue);
                        amdgpu_dm_psr_disable_all(dm);
+               }
 
                dm_enable_per_frame_crtc_master_sync(dc_state);
                mutex_lock(&dm->dc_lock);
index c6b8b835b08a7c02afdcaca829f2cd6cf4b33a08..d1d353a7c77d350767025deec33a7b90d3ed37eb 100644 (file)
@@ -91,12 +91,14 @@ struct dm_compressor_info {
  * @work: Kernel work data for the work event
  * @dm: amdgpu display manager device
  * @acrtc: amdgpu CRTC instance for which the event has occurred
+ * @stream: DC stream for which the event has occurred
  * @enable: true if enabling vblank
  */
 struct vblank_control_work {
        struct work_struct work;
        struct amdgpu_display_manager *dm;
        struct amdgpu_crtc *acrtc;
+       struct dc_stream_state *stream;
        bool enable;
 };
 
index f3b93ba69a27031eaec628321ee1859f746f0282..79b5f9999fec5eae0f2e3eab555f745968e1a0bc 100644 (file)
@@ -33,6 +33,7 @@ struct dm_irq_params {
        struct mod_vrr_params vrr_params;
        struct dc_stream_state *stream;
        int active_planes;
+       bool allow_psr_entry;
        struct mod_freesync_config freesync_config;
 
 #ifdef CONFIG_DEBUG_FS