drm/i915/dg2: Enable Wa_14014475959 - RCS / CCS context exit

There is bug in DG2 where if the CCS contexts switches out while the RCS
is running it can cause memory corruption. To workaround this add an
atomic to a memory address with a value 1 and semaphore wait to the same
address for a value of 0. The GuC firmware is responsible for writing 0
to the memory address when it is safe for the context to switch out.
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: John Harrison <John.C.Harrison@Intel.com>
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220415224025.3693037-6-umesh.nerlige.ramappa@intel.com

drivers/gpu/drm/i915/gt/gen8_engine_cs.c

@@ -583,6 +583,43 @@ static u32 *gen12_emit_preempt_busywait(struct i915_request *rq, u32 *cs)
 	return cs;
 }
 
+/* Wa_14014475959:dg2 */
+#define CCS_SEMAPHORE_PPHWSP_OFFSET	0x540
+static u32 ccs_semaphore_offset(struct i915_request *rq)
+{
+	return i915_ggtt_offset(rq->context->state) +
+		(LRC_PPHWSP_PN * PAGE_SIZE) + CCS_SEMAPHORE_PPHWSP_OFFSET;
+}
+
+/* Wa_14014475959:dg2 */
+static u32 *ccs_emit_wa_busywait(struct i915_request *rq, u32 *cs)
+{
+	int i;
+
+	*cs++ = MI_ATOMIC_INLINE | MI_ATOMIC_GLOBAL_GTT | MI_ATOMIC_CS_STALL |
+		MI_ATOMIC_MOVE;
+	*cs++ = ccs_semaphore_offset(rq);
+	*cs++ = 0;
+	*cs++ = 1;
+
+	/*
+	 * When MI_ATOMIC_INLINE_DATA set this command must be 11 DW + (1 NOP)
+	 * to align. 4 DWs above + 8 filler DWs here.
+	 */
+	for (i = 0; i < 8; ++i)
+		*cs++ = 0;
+
+	*cs++ = MI_SEMAPHORE_WAIT |
+		MI_SEMAPHORE_GLOBAL_GTT |
+		MI_SEMAPHORE_POLL |
+		MI_SEMAPHORE_SAD_EQ_SDD;
+	*cs++ = 0;
+	*cs++ = ccs_semaphore_offset(rq);
+	*cs++ = 0;
+
+	return cs;
+}
+
 static __always_inline u32*
 gen12_emit_fini_breadcrumb_tail(struct i915_request *rq, u32 *cs)
 {
@@ -593,6 +630,10 @@ gen12_emit_fini_breadcrumb_tail(struct i915_request *rq, u32 *cs)
 	    !intel_uc_uses_guc_submission(&rq->engine->gt->uc))
 		cs = gen12_emit_preempt_busywait(rq, cs);
 
+	/* Wa_14014475959:dg2 */
+	if (intel_engine_uses_wa_hold_ccs_switchout(rq->engine))
+		cs = ccs_emit_wa_busywait(rq, cs);
+
 	rq->tail = intel_ring_offset(rq, cs);
 	assert_ring_tail_valid(rq->ring, rq->tail);
 

drivers/gpu/drm/i915/gt/intel_engine_types.h

@@ -529,6 +529,7 @@ struct intel_engine_cs {
 #define I915_ENGINE_HAS_RCS_REG_STATE  BIT(9)
 #define I915_ENGINE_HAS_EU_PRIORITY    BIT(10)
 #define I915_ENGINE_FIRST_RENDER_COMPUTE BIT(11)
+#define I915_ENGINE_USES_WA_HOLD_CCS_SWITCHOUT BIT(12)
 	unsigned int flags;
 
 	/*
@@ -629,6 +630,13 @@ intel_engine_has_relative_mmio(const struct intel_engine_cs * const engine)
 	return engine->flags & I915_ENGINE_HAS_RELATIVE_MMIO;
 }
 
+/* Wa_14014475959:dg2 */
+static inline bool
+intel_engine_uses_wa_hold_ccs_switchout(struct intel_engine_cs *engine)
+{
+	return engine->flags & I915_ENGINE_USES_WA_HOLD_CCS_SWITCHOUT;
+}
+
 #define instdone_has_slice(dev_priv___, sseu___, slice___) \
 	((GRAPHICS_VER(dev_priv___) == 7 ? 1 : ((sseu___)->slice_mask)) & BIT(slice___))
 

drivers/gpu/drm/i915/gt/intel_gpu_commands.h

@@ -134,6 +134,13 @@
 #define   MI_MEM_VIRTUAL	(1 << 22) /* 945,g33,965 */
 #define   MI_USE_GGTT		(1 << 22) /* g4x+ */
 #define MI_STORE_DWORD_INDEX	MI_INSTR(0x21, 1)
+#define MI_ATOMIC		MI_INSTR(0x2f, 1)
+#define MI_ATOMIC_INLINE	(MI_INSTR(0x2f, 9) | MI_ATOMIC_INLINE_DATA)
+#define   MI_ATOMIC_GLOBAL_GTT		(1 << 22)
+#define   MI_ATOMIC_INLINE_DATA		(1 << 18)
+#define   MI_ATOMIC_CS_STALL		(1 << 17)
+#define	  MI_ATOMIC_MOVE		(0x4 << 8)
+
 /*
  * Official intel docs are somewhat sloppy concerning MI_LOAD_REGISTER_IMM:
  * - Always issue a MI_NOOP _before_ the MI_LOAD_REGISTER_IMM - otherwise hw

drivers/gpu/drm/i915/gt/uc/intel_guc.c

@@ -296,6 +296,10 @@ static u32 guc_ctl_wa_flags(struct intel_guc *guc)
 	if (IS_DG2_GRAPHICS_STEP(gt->i915, G10, STEP_A0, STEP_B0))
 		flags |= GUC_WA_GAM_CREDITS;
 
+	/* Wa_14014475959:dg2 */
+	if (IS_DG2(gt->i915))
+		flags |= GUC_WA_HOLD_CCS_SWITCHOUT;
+
 	/*
 	 * Wa_14012197797:dg2_g10:a0,dg2_g11:a0
 	 * Wa_22011391025:dg2_g10,dg2_g11,dg2_g12

drivers/gpu/drm/i915/gt/uc/intel_guc_fwif.h

@@ -102,6 +102,7 @@
 #define   GUC_WA_DUAL_QUEUE		BIT(11)
 #define   GUC_WA_RCS_RESET_BEFORE_RC6	BIT(13)
 #define   GUC_WA_PRE_PARSER		BIT(14)
+#define   GUC_WA_HOLD_CCS_SWITCHOUT	BIT(17)
 #define   GUC_WA_POLLCS			BIT(18)
 
 #define GUC_CTL_FEATURE			2

drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c

@@ -3897,6 +3897,10 @@ static void guc_default_vfuncs(struct intel_engine_cs *engine)
 	engine->flags |= I915_ENGINE_HAS_PREEMPTION;
 	engine->flags |= I915_ENGINE_HAS_TIMESLICES;
 
+	/* Wa_14014475959:dg2 */
+	if (IS_DG2(engine->i915) && engine->class == COMPUTE_CLASS)
+		engine->flags |= I915_ENGINE_USES_WA_HOLD_CCS_SWITCHOUT;
+
 	/*
 	 * TODO: GuC supports timeslicing and semaphores as well, but they're
 	 * handled by the firmware so some minor tweaks are required before