drm/amdgpu: Make SDMA phase quantum configurable

Set a configurable SDMA phase quantum when enabling SDMA context
switching. The default value significantly reduces SDMA latency
in page table updates when user-mode SDMA queues have concurrent
activity, compared to the initial HW setting.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Andres Rodriguez <andres.rodriguez@amd.com>
Reviewed-by: Shaoyun Liu <shaoyun.liu@amd.com>
Acked-by: Chunming Zhou <david1.zhou@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu.h

@@ -107,6 +107,7 @@ extern unsigned amdgpu_pcie_gen_cap;
 extern unsigned amdgpu_pcie_lane_cap;
 extern unsigned amdgpu_cg_mask;
 extern unsigned amdgpu_pg_mask;
+extern unsigned amdgpu_sdma_phase_quantum;
 extern char *amdgpu_disable_cu;
 extern char *amdgpu_virtual_display;
 extern unsigned amdgpu_pp_feature_mask;

drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c

@@ -108,6 +108,7 @@ unsigned amdgpu_pcie_gen_cap = 0;
 unsigned amdgpu_pcie_lane_cap = 0;
 unsigned amdgpu_cg_mask = 0xffffffff;
 unsigned amdgpu_pg_mask = 0xffffffff;
+unsigned amdgpu_sdma_phase_quantum = 32;
 char *amdgpu_disable_cu = NULL;
 char *amdgpu_virtual_display = NULL;
 unsigned amdgpu_pp_feature_mask = 0xffffffff;
@@ -227,6 +228,9 @@ module_param_named(cg_mask, amdgpu_cg_mask, uint, 0444);
 MODULE_PARM_DESC(pg_mask, "Powergating flags mask (0 = disable power gating)");
 module_param_named(pg_mask, amdgpu_pg_mask, uint, 0444);
 
+MODULE_PARM_DESC(sdma_phase_quantum, "SDMA context switch phase quantum (x 1K GPU clock cycles, 0 = no change (default 32))");
+module_param_named(sdma_phase_quantum, amdgpu_sdma_phase_quantum, uint, 0444);
+
 MODULE_PARM_DESC(disable_cu, "Disable CUs (se.sh.cu,...)");
 module_param_named(disable_cu, amdgpu_disable_cu, charp, 0444);
 

drivers/gpu/drm/amd/amdgpu/cik_sdma.c

@@ -351,14 +351,44 @@ static void cik_sdma_rlc_stop(struct amdgpu_device *adev)
  */
 static void cik_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
 {
-	u32 f32_cntl;
+	u32 f32_cntl, phase_quantum = 0;
 	int i;
 
+	if (amdgpu_sdma_phase_quantum) {
+		unsigned value = amdgpu_sdma_phase_quantum;
+		unsigned unit = 0;
+
+		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
+				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
+			value = (value + 1) >> 1;
+			unit++;
+		}
+		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
+			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
+			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
+				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
+			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
+				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
+			WARN_ONCE(1,
+			"clamping sdma_phase_quantum to %uK clock cycles\n",
+				  value << unit);
+		}
+		phase_quantum =
+			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
+			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
+	}
+
 	for (i = 0; i < adev->sdma.num_instances; i++) {
 		f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
 		if (enable) {
 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
 					AUTO_CTXSW_ENABLE, 1);
+			if (amdgpu_sdma_phase_quantum) {
+				WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
+				       phase_quantum);
+				WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
+				       phase_quantum);
+			}
 		} else {
 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
 					AUTO_CTXSW_ENABLE, 0);

drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c

@@ -551,9 +551,33 @@ static void sdma_v3_0_rlc_stop(struct amdgpu_device *adev)
  */
 static void sdma_v3_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
 {
-	u32 f32_cntl;
+	u32 f32_cntl, phase_quantum = 0;
 	int i;
 
+	if (amdgpu_sdma_phase_quantum) {
+		unsigned value = amdgpu_sdma_phase_quantum;
+		unsigned unit = 0;
+
+		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
+				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
+			value = (value + 1) >> 1;
+			unit++;
+		}
+		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
+			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
+			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
+				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
+			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
+				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
+			WARN_ONCE(1,
+			"clamping sdma_phase_quantum to %uK clock cycles\n",
+				  value << unit);
+		}
+		phase_quantum =
+			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
+			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
+	}
+
 	for (i = 0; i < adev->sdma.num_instances; i++) {
 		f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
 		if (enable) {
@@ -561,6 +585,12 @@ static void sdma_v3_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
 					AUTO_CTXSW_ENABLE, 1);
 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
 					ATC_L1_ENABLE, 1);
+			if (amdgpu_sdma_phase_quantum) {
+				WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
+				       phase_quantum);
+				WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
+				       phase_quantum);
+			}
 		} else {
 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
 					AUTO_CTXSW_ENABLE, 0);

drivers/gpu/drm/amd/amdgpu/sdma_v4_0.c

@@ -493,13 +493,45 @@ static void sdma_v4_0_rlc_stop(struct amdgpu_device *adev)
  */
 static void sdma_v4_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
 {
-	u32 f32_cntl;
+	u32 f32_cntl, phase_quantum = 0;
 	int i;
 
+	if (amdgpu_sdma_phase_quantum) {
+		unsigned value = amdgpu_sdma_phase_quantum;
+		unsigned unit = 0;
+
+		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
+				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
+			value = (value + 1) >> 1;
+			unit++;
+		}
+		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
+			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
+			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
+				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
+			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
+				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
+			WARN_ONCE(1,
+			"clamping sdma_phase_quantum to %uK clock cycles\n",
+				  value << unit);
+		}
+		phase_quantum =
+			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
+			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
+	}
+
 	for (i = 0; i < adev->sdma.num_instances; i++) {
 		f32_cntl = RREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_CNTL));
 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
 				AUTO_CTXSW_ENABLE, enable ? 1 : 0);
+		if (enable && amdgpu_sdma_phase_quantum) {
+			WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_PHASE0_QUANTUM),
+			       phase_quantum);
+			WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_PHASE1_QUANTUM),
+			       phase_quantum);
+			WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_PHASE2_QUANTUM),
+			       phase_quantum);
+		}
 		WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_CNTL), f32_cntl);
 	}