drm/xe/device: implement transient flush

Display surfaces can be tagged as transient by mapping it using one of
the various L3:XD PAT index modes on Xe2. The expectation is that KMD
needs to request transient data flush at the start of flip sequence to
ensure all transient data in L3 cache is flushed to memory. Add a
routine for this which we can then call from the display code.

v2: rebase(RK)
Signed-off-by: Nirmoy Das <nirmoy.das@intel.com>
Co-developed-by: Matthew Auld <matthew.auld@intel.com>
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Signed-off-by: Balasubramani Vivekanandan <balasubramani.vivekanandan@intel.com>
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Radhakrishna Sripada <radhakrishna.sripada@intel.com>
Acked-by: Lucas De Marchi <lucas.demarchi@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240430172850.1881525-18-radhakrishna.sripada@intel.com

drivers/gpu/drm/xe/regs/xe_gt_regs.h

@@ -321,6 +321,9 @@
 
 #define XE2LPM_L3SQCREG5			XE_REG_MCR(0xb658)
 
+#define XE2_TDF_CTRL				XE_REG(0xb418)
+#define   TRANSIENT_FLUSH_REQUEST		REG_BIT(0)
+
 #define XEHP_MERT_MOD_CTRL			XE_REG_MCR(0xcf28)
 #define RENDER_MOD_CTRL				XE_REG_MCR(0xcf2c)
 #define COMP_MOD_CTRL				XE_REG_MCR(0xcf30)

drivers/gpu/drm/xe/xe_device.c

@@ -708,6 +708,55 @@ void xe_device_wmb(struct xe_device *xe)
 		xe_mmio_write32(gt, SOFTWARE_FLAGS_SPR33, 0);
 }
 
+/**
+ * xe_device_td_flush() - Flush transient L3 cache entries
+ * @xe: The device
+ *
+ * Display engine has direct access to memory and is never coherent with L3/L4
+ * caches (or CPU caches), however KMD is responsible for specifically flushing
+ * transient L3 GPU cache entries prior to the flip sequence to ensure scanout
+ * can happen from such a surface without seeing corruption.
+ *
+ * Display surfaces can be tagged as transient by mapping it using one of the
+ * various L3:XD PAT index modes on Xe2.
+ *
+ * Note: On non-discrete xe2 platforms, like LNL, the entire L3 cache is flushed
+ * at the end of each submission via PIPE_CONTROL for compute/render, since SA
+ * Media is not coherent with L3 and we want to support render-vs-media
+ * usescases. For other engines like copy/blt the HW internally forces uncached
+ * behaviour, hence why we can skip the TDF on such platforms.
+ */
+void xe_device_td_flush(struct xe_device *xe)
+{
+	struct xe_gt *gt;
+	u8 id;
+
+	if (!IS_DGFX(xe) || GRAPHICS_VER(xe) < 20)
+		return;
+
+	for_each_gt(gt, xe, id) {
+		if (xe_gt_is_media_type(gt))
+			continue;
+
+		if (xe_force_wake_get(gt_to_fw(gt), XE_FW_GT))
+			return;
+
+		xe_mmio_write32(gt, XE2_TDF_CTRL, TRANSIENT_FLUSH_REQUEST);
+		/*
+		 * FIXME: We can likely do better here with our choice of
+		 * timeout. Currently we just assume the worst case, i.e. 150us,
+		 * which is believed to be sufficient to cover the worst case
+		 * scenario on current platforms if all cache entries are
+		 * transient and need to be flushed..
+		 */
+		if (xe_mmio_wait32(gt, XE2_TDF_CTRL, TRANSIENT_FLUSH_REQUEST, 0,
+				   150, NULL, false))
+			xe_gt_err_once(gt, "TD flush timeout\n");
+
+		xe_force_wake_put(gt_to_fw(gt), XE_FW_GT);
+	}
+}
+
 u32 xe_device_ccs_bytes(struct xe_device *xe, u64 size)
 {
 	return xe_device_has_flat_ccs(xe) ?

drivers/gpu/drm/xe/xe_device.h

@@ -167,4 +167,6 @@ void xe_device_snapshot_print(struct xe_device *xe, struct drm_printer *p);
 u64 xe_device_canonicalize_addr(struct xe_device *xe, u64 address);
 u64 xe_device_uncanonicalize_addr(struct xe_device *xe, u64 address);
 
+void xe_device_td_flush(struct xe_device *xe);
+
 #endif