drm/amdgpu: Make SDMA phase quantum configurableroc-1.2.0

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.

Change-Id: Id99c52e893d0358374ea9a3fbc2181f0c60b1b42
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>

4 files changed, 67 insertions, 2 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 53a2e612498d..09e4285de3eb 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -90,6 +90,7 @@ extern int amdgpu_sched_hw_submission;
 extern int amdgpu_powerplay;
 extern unsigned amdgpu_pcie_gen_cap;
 extern unsigned amdgpu_pcie_lane_cap;
+extern unsigned amdgpu_sdma_phase_quantum;
 
 #define AMDGPU_WAIT_IDLE_TIMEOUT_IN_MS	        3000
 #define AMDGPU_MAX_USEC_TIMEOUT			100000	/* 100 ms */
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index 48ce40823ca8..93e44ce78181 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
@@ -84,6 +84,7 @@ int amdgpu_sched_hw_submission = 2;
 int amdgpu_powerplay = -1;
 unsigned amdgpu_pcie_gen_cap = 0;
 unsigned amdgpu_pcie_lane_cap = 0;
+unsigned amdgpu_sdma_phase_quantum = 32;
 
 MODULE_PARM_DESC(vramlimit, "Restrict VRAM for testing, in megabytes");
 module_param_named(vramlimit, amdgpu_vram_limit, int, 0600);
@@ -174,6 +175,9 @@ module_param_named(pcie_gen_cap, amdgpu_pcie_gen_cap, uint, 0444);
 MODULE_PARM_DESC(pcie_lane_cap, "PCIE Lane Caps (0: autodetect (default))");
 module_param_named(pcie_lane_cap, amdgpu_pcie_lane_cap, 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);
+
 static const struct pci_device_id pciidlist[] = {
 #ifdef CONFIG_DRM_AMDGPU_CIK
 	/* Kaveri */
diff --git a/drivers/gpu/drm/amd/amdgpu/cik_sdma.c b/drivers/gpu/drm/amd/amdgpu/cik_sdma.c
index 2b96fa7e8961..c1035454023d 100644
--- a/drivers/gpu/drm/amd/amdgpu/cik_sdma.c
+++ b/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);
diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c b/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c
index 303f87ff9b54..3b6568506e3a 100644
--- a/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c
@@ -576,9 +576,33 @@ 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) {
@@ -586,6 +610,12 @@ 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);