drm/i915/display: perform transient flush

Perform manual transient cache flush prior to flip and at the end of
frontbuffer_flush. This is needed to ensure display engine doesn't see
garbage if the surface is L3:XD dirty.

Testcase: igt@xe-pat@display-vs-wb-transient
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Signed-off-by: Balasubramani Vivekanandan <balasubramani.vivekanandan@intel.com>
Acked-by: Nirmoy Das <nirmoy.das@intel.com>
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Radhakrishna Sripada <radhakrishna.sripada@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240430172850.1881525-19-radhakrishna.sripada@intel.com

drivers/gpu/drm/i915/display/intel_display.c

@@ -109,6 +109,7 @@
 #include "intel_sdvo.h"
 #include "intel_snps_phy.h"
 #include "intel_tc.h"
+#include "intel_tdf.h"
 #include "intel_tv.h"
 #include "intel_vblank.h"
 #include "intel_vdsc.h"
@@ -7233,6 +7234,8 @@ static void intel_atomic_commit_tail(struct intel_atomic_state *state)
 
 	intel_atomic_commit_fence_wait(state);
 
+	intel_td_flush(dev_priv);
+
 	drm_atomic_helper_wait_for_dependencies(&state->base);
 	drm_dp_mst_atomic_wait_for_dependencies(&state->base);
 	intel_atomic_global_state_wait_for_dependencies(state);

drivers/gpu/drm/i915/display/intel_frontbuffer.c

@@ -65,6 +65,7 @@
 #include "intel_fbc.h"
 #include "intel_frontbuffer.h"
 #include "intel_psr.h"
+#include "intel_tdf.h"
 
 /**
  * frontbuffer_flush - flush frontbuffer
@@ -93,6 +94,7 @@ static void frontbuffer_flush(struct drm_i915_private *i915,
 	trace_intel_frontbuffer_flush(i915, frontbuffer_bits, origin);
 
 	might_sleep();
+	intel_td_flush(i915);
 	intel_drrs_flush(i915, frontbuffer_bits);
 	intel_psr_flush(i915, frontbuffer_bits, origin);
 	intel_fbc_flush(i915, frontbuffer_bits, origin);

drivers/gpu/drm/i915/display/intel_tdf.h

+/* SPDX-License-Identifier: MIT */
+/*
+ * Copyright © 2024 Intel Corporation
+ */
+
+#ifndef __INTEL_TDF_H__
+#define __INTEL_TDF_H__
+
+/*
+ * TDF (Transient-Data-Flush) is needed for Xe2+ where special L3:XD caching can
+ * be enabled through various PAT index modes. Idea is to use this caching mode
+ * when for example rendering onto the display surface, with the promise that
+ * KMD will ensure transient cache entries are always flushed by the time we do
+ * the display flip, since display engine is never coherent with CPU/GPU caches.
+ */
+
+struct drm_i915_private;
+
+#ifdef I915
+static inline void intel_td_flush(struct drm_i915_private *i915) {}
+#else
+void intel_td_flush(struct drm_i915_private *i915);
+#endif
+
+#endif

drivers/gpu/drm/xe/Makefile

@@ -204,7 +204,8 @@ xe-$(CONFIG_DRM_XE_DISPLAY) += \
 	display/xe_dsb_buffer.o \
 	display/xe_fb_pin.o \
 	display/xe_hdcp_gsc.o \
-	display/xe_plane_initial.o
+	display/xe_plane_initial.o \
+	display/xe_tdf.o
 
 # SOC code shared with i915
 xe-$(CONFIG_DRM_XE_DISPLAY) += \

drivers/gpu/drm/xe/display/xe_tdf.c

+// SPDX-License-Identifier: MIT
+/*
+ * Copyright © 2024 Intel Corporation
+ */
+
+#include "xe_device.h"
+#include "intel_display_types.h"
+#include "intel_tdf.h"
+
+void intel_td_flush(struct drm_i915_private *i915)
+{
+	xe_device_td_flush(i915);
+}