Skip to contentSkip to conversation

drm/amdgpu: Ensure the DMA engine is deactivated during set ups

PATCH Superseded [email protected] 2 participants 8 messages Apr 28, 2022 – Apr 29, 2022

Subsystems (2)

v4 of 6 Superseded Latest is v6 →

lore
Share & tools
lore
#0 lore raw

PATCH v4 1/1 drm/amdgpu: Ensure the DMA engine is deactivated during set ups

From: Haohui Mai <[email protected]>

The patch fully deactivates the DMA engine before setting up the ring buffer to avoid potential data races and crashes.

The v4 patch addresses the comments from Lang Yu <[email protected]>.

---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 113 ++++++++++++++-----------
 1 file changed, 62 insertions(+), 51 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +62 -51
index 013d2dec81d0..b000117b77d0 100644
@@ -459,7 +459,6 @@ static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 se
 	}
 }
 
-
 /**
  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
  *
@@ -505,17 +504,21 @@ static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
 }
 
 /**
- * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
+ * sdma_v5_2_ctx_switch_enable_for_instance - start the async dma engines
+ * context switch for an instance
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs context switch.
+ * @instance_idx: the index of the SDMA instance
  *
- * Halt or unhalt the async dma engines context switch.
+ * Unhalt the async dma engines context switch.
  */
-static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_enable_for_instance(struct amdgpu_device *adev, int instance_idx)
 {
 	u32 f32_cntl, phase_quantum = 0;
-	int i;
+
+	if (WARN_ON(instance_idx >= adev->sdma.num_instances)) {
+		return;
+	}
 
 	if (amdgpu_sdma_phase_quantum) {
 		unsigned value = amdgpu_sdma_phase_quantum;
@@ -539,61 +542,71 @@ static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
 		phase_quantum =
 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
-	}
-
-	for (i = 0; i < adev->sdma.num_instances; i++) {
-		if (enable && amdgpu_sdma_phase_quantum) {
-			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
-			       phase_quantum);
-			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
-			       phase_quantum);
-			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
-			       phase_quantum);
-		}
 
-		if (!amdgpu_sriov_vf(adev)) {
-			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
-			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
-					AUTO_CTXSW_ENABLE, enable ? 1 : 0);
-			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
-		}
+		WREG32_SOC15_IP(GC,
+			sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE0_QUANTUM),
+			phase_quantum);
+		WREG32_SOC15_IP(GC,
+			sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE1_QUANTUM),
+		    phase_quantum);
+		WREG32_SOC15_IP(GC,
+			sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE2_QUANTUM),
+		    phase_quantum);
 	}
 
+	if (!amdgpu_sriov_vf(adev)) {
+		f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL));
+		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+				AUTO_CTXSW_ENABLE, 1);
+		WREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL), f32_cntl);
+	}
 }
 
 /**
- * sdma_v5_2_enable - stop the async dma engines
+ * sdma_v5_2_ctx_switch_disable_all - stop the async dma engines context switch
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs.
  *
- * Halt or unhalt the async dma engines.
+ * Halt the async dma engines context switch.
  */
-static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_disable_all(struct amdgpu_device *adev)
 {
 	u32 f32_cntl;
 	int i;
 
-	if (!enable) {
-		sdma_v5_2_gfx_stop(adev);
-		sdma_v5_2_rlc_stop(adev);
-	}
+	if (amdgpu_sriov_vf(adev))
+		return;
 
-	if (!amdgpu_sriov_vf(adev)) {
-		for (i = 0; i < adev->sdma.num_instances; i++) {
-			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
-			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
-			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
-		}
+	for (i = 0; i < adev->sdma.num_instances; i++) {
+		f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
+		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+				AUTO_CTXSW_ENABLE, 0);
+		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
 	}
 }
 
+/**
+ * sdma_v5_2_halt - stop the async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Halt the async dma engines.
+ */
+static void sdma_v5_2_halt(struct amdgpu_device *adev)
+{
+	sdma_v5_2_gfx_stop(adev);
+	sdma_v5_2_rlc_stop(adev);
+}
+
 /**
  * sdma_v5_2_gfx_resume - setup and start the async dma engines
  *
  * @adev: amdgpu_device pointer
  *
  * Set up the gfx DMA ring buffers and enable them.
+ * It assumes that the dma engine is stopped for each instance.
+ * The function enables the engine and preemptions sequentially for each instance.
+ *
  * Returns 0 for success, error for failure.
  */
 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
@@ -737,10 +750,7 @@ static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
 
 		ring->sched.ready = true;
 
-		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
-			sdma_v5_2_ctx_switch_enable(adev, true);
-			sdma_v5_2_enable(adev, true);
-		}
+		sdma_v5_2_ctx_switch_enable_for_instance(adev, i);
 
 		r = amdgpu_ring_test_ring(ring);
 		if (r) {
@@ -784,7 +794,7 @@ static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
 	int i, j;
 
 	/* halt the MEs */
-	sdma_v5_2_enable(adev, false);
+	sdma_v5_2_halt(adev);
 
 	for (i = 0; i < adev->sdma.num_instances; i++) {
 		if (!adev->sdma.instance[i].fw)
@@ -856,8 +866,8 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
 	int r = 0;
 
 	if (amdgpu_sriov_vf(adev)) {
-		sdma_v5_2_ctx_switch_enable(adev, false);
-		sdma_v5_2_enable(adev, false);
+		sdma_v5_2_ctx_switch_disable_all(adev);
+		sdma_v5_2_halt(adev);
 
 		/* set RB registers */
 		r = sdma_v5_2_gfx_resume(adev);
@@ -865,6 +875,9 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
 	}
 
 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
+		sdma_v5_2_ctx_switch_disable_all(adev);
+		sdma_v5_2_halt(adev);
+
 		r = sdma_v5_2_load_microcode(adev);
 		if (r)
 			return r;
@@ -881,12 +894,10 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
 		amdgpu_gfx_off_ctrl(adev, false);
 
 	sdma_v5_2_soft_reset(adev);
-	/* unhalt the MEs */
-	sdma_v5_2_enable(adev, true);
-	/* enable sdma ring preemption */
-	sdma_v5_2_ctx_switch_enable(adev, true);
 
-	/* start the gfx rings and rlc compute queues */
+	/* Soft reset supposes to disable the dma engine and preemption.
+	 * Now start the gfx rings and rlc compute queues.
+	 */
 	r = sdma_v5_2_gfx_resume(adev);
 	if (adev->in_s0ix)
 		amdgpu_gfx_off_ctrl(adev, true);
@@ -1340,8 +1351,8 @@ static int sdma_v5_2_hw_fini(void *handle)
 	if (amdgpu_sriov_vf(adev))
 		return 0;
 
-	sdma_v5_2_ctx_switch_enable(adev, false);
-	sdma_v5_2_enable(adev, false);
+	sdma_v5_2_ctx_switch_disable_all(adev);
+	sdma_v5_2_halt(adev);
 
 	return 0;
 }
Signature · 1 line
2.25.1
#1 lore raw
Quoted 187 lines · On 04/28/ , [email protected] wrote:

From: Haohui Mai <[email protected]>

The patch fully deactivates the DMA engine before setting up the ring buffer to avoid potential data races and crashes.

The v4 patch addresses the comments from Lang Yu <[email protected]>.

---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 113 ++++++++++++++-----------
 1 file changed, 62 insertions(+), 51 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +54 -44
index 013d2dec81d0..b000117b77d0 100644
@@ -459,7 +459,6 @@ static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 se
 	}
 }
 
-
 /**
  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
  *
@@ -505,17 +504,21 @@ static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
 }
 
 /**
- * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
+ * sdma_v5_2_ctx_switch_enable_for_instance - start the async dma engines
+ * context switch for an instance
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs context switch.
+ * @instance_idx: the index of the SDMA instance
  *
- * Halt or unhalt the async dma engines context switch.
+ * Unhalt the async dma engines context switch.
  */
-static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_enable_for_instance(struct amdgpu_device *adev, int instance_idx)
 {
 	u32 f32_cntl, phase_quantum = 0;
-	int i;
+
+	if (WARN_ON(instance_idx >= adev->sdma.num_instances)) {
+		return;
+	}
 
 	if (amdgpu_sdma_phase_quantum) {
 		unsigned value = amdgpu_sdma_phase_quantum;
@@ -539,61 +542,71 @@ static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
 		phase_quantum =
 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
-	}
-
-	for (i = 0; i < adev->sdma.num_instances; i++) {
-		if (enable && amdgpu_sdma_phase_quantum) {
-			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
-			       phase_quantum);
-			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
-			       phase_quantum);
-			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
-			       phase_quantum);
-		}
 
-		if (!amdgpu_sriov_vf(adev)) {
-			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
-			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
-					AUTO_CTXSW_ENABLE, enable ? 1 : 0);
-			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
-		}
+		WREG32_SOC15_IP(GC,
+			sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE0_QUANTUM),
+			phase_quantum);
+		WREG32_SOC15_IP(GC,
+			sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE1_QUANTUM),
+		    phase_quantum);
+		WREG32_SOC15_IP(GC,
+			sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE2_QUANTUM),
+		    phase_quantum);
 	}
 
+	if (!amdgpu_sriov_vf(adev)) {
+		f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL));
+		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+				AUTO_CTXSW_ENABLE, 1);
+		WREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL), f32_cntl);
+	}
 }
 
 /**
- * sdma_v5_2_enable - stop the async dma engines
+ * sdma_v5_2_ctx_switch_disable_all - stop the async dma engines context switch
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs.
  *
- * Halt or unhalt the async dma engines.
+ * Halt the async dma engines context switch.
  */
-static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_disable_all(struct amdgpu_device *adev)
 {
 	u32 f32_cntl;
 	int i;
 
-	if (!enable) {
-		sdma_v5_2_gfx_stop(adev);
-		sdma_v5_2_rlc_stop(adev);
-	}
+	if (amdgpu_sriov_vf(adev))
+		return;
 
-	if (!amdgpu_sriov_vf(adev)) {
-		for (i = 0; i < adev->sdma.num_instances; i++) {
-			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
-			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
-			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
-		}
+	for (i = 0; i < adev->sdma.num_instances; i++) {
+		f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
+		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+				AUTO_CTXSW_ENABLE, 0);
+		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
 	}
 }
 
+/**
+ * sdma_v5_2_halt - stop the async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Halt the async dma engines.
+ */
+static void sdma_v5_2_halt(struct amdgpu_device *adev)
+{
+	sdma_v5_2_gfx_stop(adev);
+	sdma_v5_2_rlc_stop(adev);
+}
+
 /**
  * sdma_v5_2_gfx_resume - setup and start the async dma engines
  *
  * @adev: amdgpu_device pointer
  *
  * Set up the gfx DMA ring buffers and enable them.
+ * It assumes that the dma engine is stopped for each instance.
+ * The function enables the engine and preemptions sequentially for each instance.
+ *
  * Returns 0 for success, error for failure.
  */
 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
@@ -737,10 +750,7 @@ static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
 
 		ring->sched.ready = true;
 
-		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
-			sdma_v5_2_ctx_switch_enable(adev, true);
-			sdma_v5_2_enable(adev, true);
-		}
+		sdma_v5_2_ctx_switch_enable_for_instance(adev, i);
 
 		r = amdgpu_ring_test_ring(ring);
 		if (r) {
@@ -784,7 +794,7 @@ static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
 	int i, j;
 
 	/* halt the MEs */
-	sdma_v5_2_enable(adev, false);
+	sdma_v5_2_halt(adev);
 
 	for (i = 0; i < adev->sdma.num_instances; i++) {
 		if (!adev->sdma.instance[i].fw)
@@ -856,8 +866,8 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
 	int r = 0;
 
 	if (amdgpu_sriov_vf(adev)) {
-		sdma_v5_2_ctx_switch_enable(adev, false);
-		sdma_v5_2_enable(adev, false);
+		sdma_v5_2_ctx_switch_disable_all(adev);
+		sdma_v5_2_halt(adev);
 
 		/* set RB registers */
 		r = sdma_v5_2_gfx_resume(adev);

drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c static int sdma_v5_2_start(struct amdgpu_device *adev) · line 879

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +2 -0
@@ -865,6 +875,9 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
 	}
 
 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
+		sdma_v5_2_ctx_switch_disable_all(adev);
+		sdma_v5_2_halt(adev);

You don't need call these functions here.

In original code, you can see sdma_v5_2_load_microcode() calls sdma_v5_2_enable(adev, false) to halt the engine before load ucode.

What I mean is sdma_v5_2_halt(adev) should have no functional change with original sdma_v5_2_enable(adev, false), like this,

static void sdma_v5_2_halt(struct amdgpu_device *adev) {
	u32 f32_cntl;
	int i;

	sdma_v5_2_gfx_stop(adev);
	sdma_v5_2_rlc_stop(adev);
	
	if (!amdgpu_sriov_vf(adev)) {
		for (i = 0; i < adev->sdma.num_instances; i++) {
			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
		}
	}	
}

But you remove the codes to set the HLAT bit of register mmSDMA0_F32_CNTL.

Regards,
Lang

Quoted 33 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +5 -7
 		r = sdma_v5_2_load_microcode(adev);
 		if (r)
 			return r;
@@ -881,12 +894,10 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
 		amdgpu_gfx_off_ctrl(adev, false);
 
 	sdma_v5_2_soft_reset(adev);
-	/* unhalt the MEs */
-	sdma_v5_2_enable(adev, true);
-	/* enable sdma ring preemption */
-	sdma_v5_2_ctx_switch_enable(adev, true);
 
-	/* start the gfx rings and rlc compute queues */
+	/* Soft reset supposes to disable the dma engine and preemption.
+	 * Now start the gfx rings and rlc compute queues.
+	 */
 	r = sdma_v5_2_gfx_resume(adev);
 	if (adev->in_s0ix)
 		amdgpu_gfx_off_ctrl(adev, true);
@@ -1340,8 +1351,8 @@ static int sdma_v5_2_hw_fini(void *handle)
 	if (amdgpu_sriov_vf(adev))
 		return 0;
 
-	sdma_v5_2_ctx_switch_enable(adev, false);
-	sdma_v5_2_enable(adev, false);
+	sdma_v5_2_ctx_switch_disable_all(adev);
+	sdma_v5_2_halt(adev);
 
 	return 0;
 }
Signature · 1 line
2.25.1
author
#2 lore raw

If I understand correctly, the original code will disable the HALT bit of the register mmSDMA0_F32_CNTL twice on other code paths -- one in the sdma_v5_2_ctx_switch_enable() and the other one in sdma_v5_2_enable().

The change ensures that the bit is only disabled once. Just wondering, which one is the expected behavior?

~Haohui

Quoted 257 lines · Lang Yu
Quoted 187 lines · On 04/28/ , [email protected] wrote:

From: Haohui Mai <[email protected]>

The patch fully deactivates the DMA engine before setting up the ring buffer to avoid potential data races and crashes.

The v4 patch addresses the comments from Lang Yu <[email protected]>.

---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 113 ++++++++++++++-----------
 1 file changed, 62 insertions(+), 51 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +54 -44
index 013d2dec81d0..b000117b77d0 100644
@@ -459,7 +459,6 @@ static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 se
      }
 }
 
-
 /**
  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
  *
@@ -505,17 +504,21 @@ static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
 }
 
 /**
- * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
+ * sdma_v5_2_ctx_switch_enable_for_instance - start the async dma engines
+ * context switch for an instance
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs context switch.
+ * @instance_idx: the index of the SDMA instance
  *
- * Halt or unhalt the async dma engines context switch.
+ * Unhalt the async dma engines context switch.
  */
-static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_enable_for_instance(struct amdgpu_device *adev, int instance_idx)
 {
      u32 f32_cntl, phase_quantum = 0;
-     int i;
+
+     if (WARN_ON(instance_idx >= adev->sdma.num_instances)) {
+             return;
+     }
 
      if (amdgpu_sdma_phase_quantum) {
              unsigned value = amdgpu_sdma_phase_quantum;
@@ -539,61 +542,71 @@ static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
              phase_quantum =
                      value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
                      unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
-     }
-
-     for (i = 0; i < adev->sdma.num_instances; i++) {
-             if (enable && amdgpu_sdma_phase_quantum) {
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
-                            phase_quantum);
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
-                            phase_quantum);
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
-                            phase_quantum);
-             }
 
-             if (!amdgpu_sriov_vf(adev)) {
-                     f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
-                     f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
-                                     AUTO_CTXSW_ENABLE, enable ? 1 : 0);
-                     WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
-             }
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE0_QUANTUM),
+                     phase_quantum);
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE1_QUANTUM),
+                 phase_quantum);
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE2_QUANTUM),
+                 phase_quantum);
      }
 
+     if (!amdgpu_sriov_vf(adev)) {
+             f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL));
+             f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+                             AUTO_CTXSW_ENABLE, 1);
+             WREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL), f32_cntl);
+     }
 }
 
 /**
- * sdma_v5_2_enable - stop the async dma engines
+ * sdma_v5_2_ctx_switch_disable_all - stop the async dma engines context switch
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs.
  *
- * Halt or unhalt the async dma engines.
+ * Halt the async dma engines context switch.
  */
-static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_disable_all(struct amdgpu_device *adev)
 {
      u32 f32_cntl;
      int i;
 
-     if (!enable) {
-             sdma_v5_2_gfx_stop(adev);
-             sdma_v5_2_rlc_stop(adev);
-     }
+     if (amdgpu_sriov_vf(adev))
+             return;
 
-     if (!amdgpu_sriov_vf(adev)) {
-             for (i = 0; i < adev->sdma.num_instances; i++) {
-                     f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
-                     f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
-                     WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
-             }
+     for (i = 0; i < adev->sdma.num_instances; i++) {
+             f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
+             f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+                             AUTO_CTXSW_ENABLE, 0);
+             WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
      }
 }
 
+/**
+ * sdma_v5_2_halt - stop the async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Halt the async dma engines.
+ */
+static void sdma_v5_2_halt(struct amdgpu_device *adev)
+{
+     sdma_v5_2_gfx_stop(adev);
+     sdma_v5_2_rlc_stop(adev);
+}
+
 /**
  * sdma_v5_2_gfx_resume - setup and start the async dma engines
  *
  * @adev: amdgpu_device pointer
  *
  * Set up the gfx DMA ring buffers and enable them.
+ * It assumes that the dma engine is stopped for each instance.
+ * The function enables the engine and preemptions sequentially for each instance.
+ *
  * Returns 0 for success, error for failure.
  */
 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
@@ -737,10 +750,7 @@ static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
 
              ring->sched.ready = true;
 
-             if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
-                     sdma_v5_2_ctx_switch_enable(adev, true);
-                     sdma_v5_2_enable(adev, true);
-             }
+             sdma_v5_2_ctx_switch_enable_for_instance(adev, i);
 
              r = amdgpu_ring_test_ring(ring);
              if (r) {
@@ -784,7 +794,7 @@ static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
      int i, j;
 
      /* halt the MEs */
-     sdma_v5_2_enable(adev, false);
+     sdma_v5_2_halt(adev);
 
      for (i = 0; i < adev->sdma.num_instances; i++) {
              if (!adev->sdma.instance[i].fw)
@@ -856,8 +866,8 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
      int r = 0;
 
      if (amdgpu_sriov_vf(adev)) {
-             sdma_v5_2_ctx_switch_enable(adev, false);
-             sdma_v5_2_enable(adev, false);
+             sdma_v5_2_ctx_switch_disable_all(adev);
+             sdma_v5_2_halt(adev);
 
              /* set RB registers */
              r = sdma_v5_2_gfx_resume(adev);

drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c static int sdma_v5_2_start(struct amdgpu_device *adev) · line 879

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +2 -0
@@ -865,6 +875,9 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
      }
 
      if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
+             sdma_v5_2_ctx_switch_disable_all(adev);
+             sdma_v5_2_halt(adev);

You don't need call these functions here.

In original code, you can see sdma_v5_2_load_microcode() calls sdma_v5_2_enable(adev, false) to halt the engine before load ucode.

What I mean is sdma_v5_2_halt(adev) should have no functional change with original sdma_v5_2_enable(adev, false), like this,

static void sdma_v5_2_halt(struct amdgpu_device *adev) {
        u32 f32_cntl;
        int i;

        sdma_v5_2_gfx_stop(adev);
        sdma_v5_2_rlc_stop(adev);

        if (!amdgpu_sriov_vf(adev)) {
                for (i = 0; i < adev->sdma.num_instances; i++) {
                        f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
                        f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
                        WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
                }
        }
}

But you remove the codes to set the HLAT bit of register mmSDMA0_F32_CNTL.

Regards,
Lang

Quoted 33 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +5 -8
              r = sdma_v5_2_load_microcode(adev);
              if (r)
                      return r;
@@ -881,12 +894,10 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
              amdgpu_gfx_off_ctrl(adev, false);
 
      sdma_v5_2_soft_reset(adev);
-     /* unhalt the MEs */
-     sdma_v5_2_enable(adev, true);
-     /* enable sdma ring preemption */
-     sdma_v5_2_ctx_switch_enable(adev, true);
 
-     /* start the gfx rings and rlc compute queues */
+     /* Soft reset supposes to disable the dma engine and preemption.
+      * Now start the gfx rings and rlc compute queues.
+      */
      r = sdma_v5_2_gfx_resume(adev);
      if (adev->in_s0ix)
              amdgpu_gfx_off_ctrl(adev, true);
@@ -1340,8 +1351,8 @@ static int sdma_v5_2_hw_fini(void *handle)
      if (amdgpu_sriov_vf(adev))
              return 0;
 
-     sdma_v5_2_ctx_switch_enable(adev, false);
-     sdma_v5_2_enable(adev, false);
+     sdma_v5_2_ctx_switch_disable_all(adev);
+     sdma_v5_2_halt(adev);
 
      return 0;
 }
--

2.25.1

#3 lore raw
Quoted 9 lines · Haohui Mai

If I understand correctly, the original code will disable the HALT bit of the register mmSDMA0_F32_CNTL twice on other code paths -- one in the sdma_v5_2_ctx_switch_enable() and the other one in sdma_v5_2_enable().

The change ensures that the bit is only disabled once. Just wondering, which one is the expected behavior?

~Haohui

The HALT bit of the register mmSDMA0_F32_CNTL is not set/reset in original sdma_v5_2_ctx_switch_enable()
(https://gitlab.freedesktop.org/agd5f/linux/-/blob/amd-staging-drm-next/drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c#L523)

Regards,
Lang

Quoted 258 lines · Lang Yu
Quoted 257 lines · On 04/28/ , [email protected] wrote:
Quoted 187 lines

From: Haohui Mai <[email protected]>

The patch fully deactivates the DMA engine before setting up the ring buffer to avoid potential data races and crashes.

The v4 patch addresses the comments from Lang Yu <[email protected]>.

---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 113 ++++++++++++++-----------
 1 file changed, 62 insertions(+), 51 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +54 -44
index 013d2dec81d0..b000117b77d0 100644
@@ -459,7 +459,6 @@ static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 se
      }
 }
 
-
 /**
  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
  *
@@ -505,17 +504,21 @@ static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
 }
 
 /**
- * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
+ * sdma_v5_2_ctx_switch_enable_for_instance - start the async dma engines
+ * context switch for an instance
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs context switch.
+ * @instance_idx: the index of the SDMA instance
  *
- * Halt or unhalt the async dma engines context switch.
+ * Unhalt the async dma engines context switch.
  */
-static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_enable_for_instance(struct amdgpu_device *adev, int instance_idx)
 {
      u32 f32_cntl, phase_quantum = 0;
-     int i;
+
+     if (WARN_ON(instance_idx >= adev->sdma.num_instances)) {
+             return;
+     }
 
      if (amdgpu_sdma_phase_quantum) {
              unsigned value = amdgpu_sdma_phase_quantum;
@@ -539,61 +542,71 @@ static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
              phase_quantum =
                      value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
                      unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
-     }
-
-     for (i = 0; i < adev->sdma.num_instances; i++) {
-             if (enable && amdgpu_sdma_phase_quantum) {
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
-                            phase_quantum);
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
-                            phase_quantum);
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
-                            phase_quantum);
-             }
 
-             if (!amdgpu_sriov_vf(adev)) {
-                     f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
-                     f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
-                                     AUTO_CTXSW_ENABLE, enable ? 1 : 0);
-                     WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
-             }
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE0_QUANTUM),
+                     phase_quantum);
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE1_QUANTUM),
+                 phase_quantum);
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE2_QUANTUM),
+                 phase_quantum);
      }
 
+     if (!amdgpu_sriov_vf(adev)) {
+             f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL));
+             f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+                             AUTO_CTXSW_ENABLE, 1);
+             WREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL), f32_cntl);
+     }
 }
 
 /**
- * sdma_v5_2_enable - stop the async dma engines
+ * sdma_v5_2_ctx_switch_disable_all - stop the async dma engines context switch
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs.
  *
- * Halt or unhalt the async dma engines.
+ * Halt the async dma engines context switch.
  */
-static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_disable_all(struct amdgpu_device *adev)
 {
      u32 f32_cntl;
      int i;
 
-     if (!enable) {
-             sdma_v5_2_gfx_stop(adev);
-             sdma_v5_2_rlc_stop(adev);
-     }
+     if (amdgpu_sriov_vf(adev))
+             return;
 
-     if (!amdgpu_sriov_vf(adev)) {
-             for (i = 0; i < adev->sdma.num_instances; i++) {
-                     f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
-                     f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
-                     WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
-             }
+     for (i = 0; i < adev->sdma.num_instances; i++) {
+             f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
+             f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+                             AUTO_CTXSW_ENABLE, 0);
+             WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
      }
 }
 
+/**
+ * sdma_v5_2_halt - stop the async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Halt the async dma engines.
+ */
+static void sdma_v5_2_halt(struct amdgpu_device *adev)
+{
+     sdma_v5_2_gfx_stop(adev);
+     sdma_v5_2_rlc_stop(adev);
+}
+
 /**
  * sdma_v5_2_gfx_resume - setup and start the async dma engines
  *
  * @adev: amdgpu_device pointer
  *
  * Set up the gfx DMA ring buffers and enable them.
+ * It assumes that the dma engine is stopped for each instance.
+ * The function enables the engine and preemptions sequentially for each instance.
+ *
  * Returns 0 for success, error for failure.
  */
 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
@@ -737,10 +750,7 @@ static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
 
              ring->sched.ready = true;
 
-             if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
-                     sdma_v5_2_ctx_switch_enable(adev, true);
-                     sdma_v5_2_enable(adev, true);
-             }
+             sdma_v5_2_ctx_switch_enable_for_instance(adev, i);
 
              r = amdgpu_ring_test_ring(ring);
              if (r) {
@@ -784,7 +794,7 @@ static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
      int i, j;
 
      /* halt the MEs */
-     sdma_v5_2_enable(adev, false);
+     sdma_v5_2_halt(adev);
 
      for (i = 0; i < adev->sdma.num_instances; i++) {
              if (!adev->sdma.instance[i].fw)
@@ -856,8 +866,8 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
      int r = 0;
 
      if (amdgpu_sriov_vf(adev)) {
-             sdma_v5_2_ctx_switch_enable(adev, false);
-             sdma_v5_2_enable(adev, false);
+             sdma_v5_2_ctx_switch_disable_all(adev);
+             sdma_v5_2_halt(adev);
 
              /* set RB registers */
              r = sdma_v5_2_gfx_resume(adev);

drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c static int sdma_v5_2_start(struct amdgpu_device *adev) · line 879

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +2 -0
@@ -865,6 +875,9 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
      }
 
      if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
+             sdma_v5_2_ctx_switch_disable_all(adev);
+             sdma_v5_2_halt(adev);

You don't need call these functions here.

In original code, you can see sdma_v5_2_load_microcode() calls sdma_v5_2_enable(adev, false) to halt the engine before load ucode.

What I mean is sdma_v5_2_halt(adev) should have no functional change with original sdma_v5_2_enable(adev, false), like this,

static void sdma_v5_2_halt(struct amdgpu_device *adev) {
        u32 f32_cntl;
        int i;

        sdma_v5_2_gfx_stop(adev);
        sdma_v5_2_rlc_stop(adev);

        if (!amdgpu_sriov_vf(adev)) {
                for (i = 0; i < adev->sdma.num_instances; i++) {
                        f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
                        f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
                        WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
                }
        }
}

But you remove the codes to set the HLAT bit of register mmSDMA0_F32_CNTL.

Regards,
Lang

Quoted 33 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +5 -8
              r = sdma_v5_2_load_microcode(adev);
              if (r)
                      return r;
@@ -881,12 +894,10 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
              amdgpu_gfx_off_ctrl(adev, false);
 
      sdma_v5_2_soft_reset(adev);
-     /* unhalt the MEs */
-     sdma_v5_2_enable(adev, true);
-     /* enable sdma ring preemption */
-     sdma_v5_2_ctx_switch_enable(adev, true);
 
-     /* start the gfx rings and rlc compute queues */
+     /* Soft reset supposes to disable the dma engine and preemption.
+      * Now start the gfx rings and rlc compute queues.
+      */
      r = sdma_v5_2_gfx_resume(adev);
      if (adev->in_s0ix)
              amdgpu_gfx_off_ctrl(adev, true);
@@ -1340,8 +1351,8 @@ static int sdma_v5_2_hw_fini(void *handle)
      if (amdgpu_sriov_vf(adev))
              return 0;
 
-     sdma_v5_2_ctx_switch_enable(adev, false);
-     sdma_v5_2_enable(adev, false);
+     sdma_v5_2_ctx_switch_disable_all(adev);
+     sdma_v5_2_halt(adev);
 
      return 0;
 }
--

2.25.1

author
#4 lore raw

Thanks for pointing it out. The v5 patch added the code back.

~Haohui

Quoted 276 lines · Lang Yu
Quoted 9 lines · Haohui Mai

If I understand correctly, the original code will disable the HALT bit of the register mmSDMA0_F32_CNTL twice on other code paths -- one in the sdma_v5_2_ctx_switch_enable() and the other one in sdma_v5_2_enable().

The change ensures that the bit is only disabled once. Just wondering, which one is the expected behavior?

~Haohui

The HALT bit of the register mmSDMA0_F32_CNTL is not set/reset in original sdma_v5_2_ctx_switch_enable()
(https://gitlab.freedesktop.org/agd5f/linux/-/blob/amd-staging-drm-next/drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c#L523)

Regards,
Lang

Quoted 258 lines · Lang Yu
Quoted 257 lines · On 04/28/ , [email protected] wrote:
Quoted 187 lines

From: Haohui Mai <[email protected]>

The patch fully deactivates the DMA engine before setting up the ring buffer to avoid potential data races and crashes.

The v4 patch addresses the comments from Lang Yu <[email protected]>.

---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 113 ++++++++++++++-----------
 1 file changed, 62 insertions(+), 51 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +54 -44
index 013d2dec81d0..b000117b77d0 100644
@@ -459,7 +459,6 @@ static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 se
      }
 }
 
-
 /**
  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
  *
@@ -505,17 +504,21 @@ static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
 }
 
 /**
- * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
+ * sdma_v5_2_ctx_switch_enable_for_instance - start the async dma engines
+ * context switch for an instance
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs context switch.
+ * @instance_idx: the index of the SDMA instance
  *
- * Halt or unhalt the async dma engines context switch.
+ * Unhalt the async dma engines context switch.
  */
-static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_enable_for_instance(struct amdgpu_device *adev, int instance_idx)
 {
      u32 f32_cntl, phase_quantum = 0;
-     int i;
+
+     if (WARN_ON(instance_idx >= adev->sdma.num_instances)) {
+             return;
+     }
 
      if (amdgpu_sdma_phase_quantum) {
              unsigned value = amdgpu_sdma_phase_quantum;
@@ -539,61 +542,71 @@ static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
              phase_quantum =
                      value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
                      unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
-     }
-
-     for (i = 0; i < adev->sdma.num_instances; i++) {
-             if (enable && amdgpu_sdma_phase_quantum) {
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
-                            phase_quantum);
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
-                            phase_quantum);
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
-                            phase_quantum);
-             }
 
-             if (!amdgpu_sriov_vf(adev)) {
-                     f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
-                     f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
-                                     AUTO_CTXSW_ENABLE, enable ? 1 : 0);
-                     WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
-             }
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE0_QUANTUM),
+                     phase_quantum);
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE1_QUANTUM),
+                 phase_quantum);
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE2_QUANTUM),
+                 phase_quantum);
      }
 
+     if (!amdgpu_sriov_vf(adev)) {
+             f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL));
+             f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+                             AUTO_CTXSW_ENABLE, 1);
+             WREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL), f32_cntl);
+     }
 }
 
 /**
- * sdma_v5_2_enable - stop the async dma engines
+ * sdma_v5_2_ctx_switch_disable_all - stop the async dma engines context switch
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs.
  *
- * Halt or unhalt the async dma engines.
+ * Halt the async dma engines context switch.
  */
-static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_disable_all(struct amdgpu_device *adev)
 {
      u32 f32_cntl;
      int i;
 
-     if (!enable) {
-             sdma_v5_2_gfx_stop(adev);
-             sdma_v5_2_rlc_stop(adev);
-     }
+     if (amdgpu_sriov_vf(adev))
+             return;
 
-     if (!amdgpu_sriov_vf(adev)) {
-             for (i = 0; i < adev->sdma.num_instances; i++) {
-                     f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
-                     f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
-                     WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
-             }
+     for (i = 0; i < adev->sdma.num_instances; i++) {
+             f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
+             f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+                             AUTO_CTXSW_ENABLE, 0);
+             WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
      }
 }
 
+/**
+ * sdma_v5_2_halt - stop the async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Halt the async dma engines.
+ */
+static void sdma_v5_2_halt(struct amdgpu_device *adev)
+{
+     sdma_v5_2_gfx_stop(adev);
+     sdma_v5_2_rlc_stop(adev);
+}
+
 /**
  * sdma_v5_2_gfx_resume - setup and start the async dma engines
  *
  * @adev: amdgpu_device pointer
  *
  * Set up the gfx DMA ring buffers and enable them.
+ * It assumes that the dma engine is stopped for each instance.
+ * The function enables the engine and preemptions sequentially for each instance.
+ *
  * Returns 0 for success, error for failure.
  */
 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
@@ -737,10 +750,7 @@ static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
 
              ring->sched.ready = true;
 
-             if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
-                     sdma_v5_2_ctx_switch_enable(adev, true);
-                     sdma_v5_2_enable(adev, true);
-             }
+             sdma_v5_2_ctx_switch_enable_for_instance(adev, i);
 
              r = amdgpu_ring_test_ring(ring);
              if (r) {
@@ -784,7 +794,7 @@ static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
      int i, j;
 
      /* halt the MEs */
-     sdma_v5_2_enable(adev, false);
+     sdma_v5_2_halt(adev);
 
      for (i = 0; i < adev->sdma.num_instances; i++) {
              if (!adev->sdma.instance[i].fw)
@@ -856,8 +866,8 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
      int r = 0;
 
      if (amdgpu_sriov_vf(adev)) {
-             sdma_v5_2_ctx_switch_enable(adev, false);
-             sdma_v5_2_enable(adev, false);
+             sdma_v5_2_ctx_switch_disable_all(adev);
+             sdma_v5_2_halt(adev);
 
              /* set RB registers */
              r = sdma_v5_2_gfx_resume(adev);

drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c static int sdma_v5_2_start(struct amdgpu_device *adev) · line 879

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +2 -0
@@ -865,6 +875,9 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
      }
 
      if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
+             sdma_v5_2_ctx_switch_disable_all(adev);
+             sdma_v5_2_halt(adev);

You don't need call these functions here.

In original code, you can see sdma_v5_2_load_microcode() calls sdma_v5_2_enable(adev, false) to halt the engine before load ucode.

What I mean is sdma_v5_2_halt(adev) should have no functional change with original sdma_v5_2_enable(adev, false), like this,

static void sdma_v5_2_halt(struct amdgpu_device *adev) {
        u32 f32_cntl;
        int i;

        sdma_v5_2_gfx_stop(adev);
        sdma_v5_2_rlc_stop(adev);

        if (!amdgpu_sriov_vf(adev)) {
                for (i = 0; i < adev->sdma.num_instances; i++) {
                        f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
                        f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
                        WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
                }
        }
}

But you remove the codes to set the HLAT bit of register mmSDMA0_F32_CNTL.

Regards,
Lang

Quoted 33 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +5 -8
              r = sdma_v5_2_load_microcode(adev);
              if (r)
                      return r;
@@ -881,12 +894,10 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
              amdgpu_gfx_off_ctrl(adev, false);
 
      sdma_v5_2_soft_reset(adev);
-     /* unhalt the MEs */
-     sdma_v5_2_enable(adev, true);
-     /* enable sdma ring preemption */
-     sdma_v5_2_ctx_switch_enable(adev, true);
 
-     /* start the gfx rings and rlc compute queues */
+     /* Soft reset supposes to disable the dma engine and preemption.
+      * Now start the gfx rings and rlc compute queues.
+      */
      r = sdma_v5_2_gfx_resume(adev);
      if (adev->in_s0ix)
              amdgpu_gfx_off_ctrl(adev, true);
@@ -1340,8 +1351,8 @@ static int sdma_v5_2_hw_fini(void *handle)
      if (amdgpu_sriov_vf(adev))
              return 0;
 
-     sdma_v5_2_ctx_switch_enable(adev, false);
-     sdma_v5_2_enable(adev, false);
+     sdma_v5_2_ctx_switch_disable_all(adev);
+     sdma_v5_2_halt(adev);
 
      return 0;
 }
--

2.25.1

#5 lore raw
Quoted 3 lines · Haohui Mai

Thanks for pointing it out. The v5 patch added the code back.

~Haohui

But you didn't remove the following codes added in v4.

if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
+ sdma_v5_2_ctx_switch_disable_all(adev);
+ sdma_v5_2_halt(adev);
+

sdma_v5_2_halt() has been called in sdma_v5_2_load_microcode().

No need to call these two functions here.

Regards,
Lang

Quoted 277 lines · Lang Yu
Quoted 276 lines · Haohui Mai
Quoted 9 lines

If I understand correctly, the original code will disable the HALT bit of the register mmSDMA0_F32_CNTL twice on other code paths -- one in the sdma_v5_2_ctx_switch_enable() and the other one in sdma_v5_2_enable().

The change ensures that the bit is only disabled once. Just wondering, which one is the expected behavior?

~Haohui

The HALT bit of the register mmSDMA0_F32_CNTL is not set/reset in original sdma_v5_2_ctx_switch_enable()
(https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgitlab.freedesktop.org%2Fagd5f%2Flinux%2F-%2Fblob%2Famd-staging-drm-next%2Fdrivers%2Fgpu%2Fdrm%2Famd%2Famdgpu%2Fsdma_v5_2.c%23L523&amp;data=05%7C01%7CLang.Yu%40amd.com%7C2df035ef6ad5411215a208da299c0253%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637868047861293833%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&amp;sdata=%2FLpd6TUyxSe96rEzbyvV0zXNPZ%2BIokiT8KtGAKea8Co%3D&amp;reserved=0)

Regards,
Lang

Quoted 258 lines · Lang Yu
Quoted 257 lines · On 04/28/ , [email protected] wrote:
Quoted 187 lines

From: Haohui Mai <[email protected]>

The patch fully deactivates the DMA engine before setting up the ring buffer to avoid potential data races and crashes.

The v4 patch addresses the comments from Lang Yu <[email protected]>.

---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 113 ++++++++++++++-----------
 1 file changed, 62 insertions(+), 51 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +54 -44
index 013d2dec81d0..b000117b77d0 100644
@@ -459,7 +459,6 @@ static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 se
      }
 }
 
-
 /**
  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
  *
@@ -505,17 +504,21 @@ static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
 }
 
 /**
- * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
+ * sdma_v5_2_ctx_switch_enable_for_instance - start the async dma engines
+ * context switch for an instance
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs context switch.
+ * @instance_idx: the index of the SDMA instance
  *
- * Halt or unhalt the async dma engines context switch.
+ * Unhalt the async dma engines context switch.
  */
-static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_enable_for_instance(struct amdgpu_device *adev, int instance_idx)
 {
      u32 f32_cntl, phase_quantum = 0;
-     int i;
+
+     if (WARN_ON(instance_idx >= adev->sdma.num_instances)) {
+             return;
+     }
 
      if (amdgpu_sdma_phase_quantum) {
              unsigned value = amdgpu_sdma_phase_quantum;
@@ -539,61 +542,71 @@ static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
              phase_quantum =
                      value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
                      unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
-     }
-
-     for (i = 0; i < adev->sdma.num_instances; i++) {
-             if (enable && amdgpu_sdma_phase_quantum) {
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
-                            phase_quantum);
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
-                            phase_quantum);
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
-                            phase_quantum);
-             }
 
-             if (!amdgpu_sriov_vf(adev)) {
-                     f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
-                     f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
-                                     AUTO_CTXSW_ENABLE, enable ? 1 : 0);
-                     WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
-             }
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE0_QUANTUM),
+                     phase_quantum);
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE1_QUANTUM),
+                 phase_quantum);
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE2_QUANTUM),
+                 phase_quantum);
      }
 
+     if (!amdgpu_sriov_vf(adev)) {
+             f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL));
+             f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+                             AUTO_CTXSW_ENABLE, 1);
+             WREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL), f32_cntl);
+     }
 }
 
 /**
- * sdma_v5_2_enable - stop the async dma engines
+ * sdma_v5_2_ctx_switch_disable_all - stop the async dma engines context switch
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs.
  *
- * Halt or unhalt the async dma engines.
+ * Halt the async dma engines context switch.
  */
-static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_disable_all(struct amdgpu_device *adev)
 {
      u32 f32_cntl;
      int i;
 
-     if (!enable) {
-             sdma_v5_2_gfx_stop(adev);
-             sdma_v5_2_rlc_stop(adev);
-     }
+     if (amdgpu_sriov_vf(adev))
+             return;
 
-     if (!amdgpu_sriov_vf(adev)) {
-             for (i = 0; i < adev->sdma.num_instances; i++) {
-                     f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
-                     f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
-                     WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
-             }
+     for (i = 0; i < adev->sdma.num_instances; i++) {
+             f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
+             f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+                             AUTO_CTXSW_ENABLE, 0);
+             WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
      }
 }
 
+/**
+ * sdma_v5_2_halt - stop the async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Halt the async dma engines.
+ */
+static void sdma_v5_2_halt(struct amdgpu_device *adev)
+{
+     sdma_v5_2_gfx_stop(adev);
+     sdma_v5_2_rlc_stop(adev);
+}
+
 /**
  * sdma_v5_2_gfx_resume - setup and start the async dma engines
  *
  * @adev: amdgpu_device pointer
  *
  * Set up the gfx DMA ring buffers and enable them.
+ * It assumes that the dma engine is stopped for each instance.
+ * The function enables the engine and preemptions sequentially for each instance.
+ *
  * Returns 0 for success, error for failure.
  */
 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
@@ -737,10 +750,7 @@ static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
 
              ring->sched.ready = true;
 
-             if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
-                     sdma_v5_2_ctx_switch_enable(adev, true);
-                     sdma_v5_2_enable(adev, true);
-             }
+             sdma_v5_2_ctx_switch_enable_for_instance(adev, i);
 
              r = amdgpu_ring_test_ring(ring);
              if (r) {
@@ -784,7 +794,7 @@ static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
      int i, j;
 
      /* halt the MEs */
-     sdma_v5_2_enable(adev, false);
+     sdma_v5_2_halt(adev);
 
      for (i = 0; i < adev->sdma.num_instances; i++) {
              if (!adev->sdma.instance[i].fw)
@@ -856,8 +866,8 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
      int r = 0;
 
      if (amdgpu_sriov_vf(adev)) {
-             sdma_v5_2_ctx_switch_enable(adev, false);
-             sdma_v5_2_enable(adev, false);
+             sdma_v5_2_ctx_switch_disable_all(adev);
+             sdma_v5_2_halt(adev);
 
              /* set RB registers */
              r = sdma_v5_2_gfx_resume(adev);

drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c static int sdma_v5_2_start(struct amdgpu_device *adev) · line 879

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +2 -0
@@ -865,6 +875,9 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
      }
 
      if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
+             sdma_v5_2_ctx_switch_disable_all(adev);
+             sdma_v5_2_halt(adev);

You don't need call these functions here.

In original code, you can see sdma_v5_2_load_microcode() calls sdma_v5_2_enable(adev, false) to halt the engine before load ucode.

What I mean is sdma_v5_2_halt(adev) should have no functional change with original sdma_v5_2_enable(adev, false), like this,

static void sdma_v5_2_halt(struct amdgpu_device *adev) {
        u32 f32_cntl;
        int i;

        sdma_v5_2_gfx_stop(adev);
        sdma_v5_2_rlc_stop(adev);

        if (!amdgpu_sriov_vf(adev)) {
                for (i = 0; i < adev->sdma.num_instances; i++) {
                        f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
                        f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
                        WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
                }
        }
}

But you remove the codes to set the HLAT bit of register mmSDMA0_F32_CNTL.

Regards,
Lang

Quoted 33 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +5 -8
              r = sdma_v5_2_load_microcode(adev);
              if (r)
                      return r;
@@ -881,12 +894,10 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
              amdgpu_gfx_off_ctrl(adev, false);
 
      sdma_v5_2_soft_reset(adev);
-     /* unhalt the MEs */
-     sdma_v5_2_enable(adev, true);
-     /* enable sdma ring preemption */
-     sdma_v5_2_ctx_switch_enable(adev, true);
 
-     /* start the gfx rings and rlc compute queues */
+     /* Soft reset supposes to disable the dma engine and preemption.
+      * Now start the gfx rings and rlc compute queues.
+      */
      r = sdma_v5_2_gfx_resume(adev);
      if (adev->in_s0ix)
              amdgpu_gfx_off_ctrl(adev, true);
@@ -1340,8 +1351,8 @@ static int sdma_v5_2_hw_fini(void *handle)
      if (amdgpu_sriov_vf(adev))
              return 0;
 
-     sdma_v5_2_ctx_switch_enable(adev, false);
-     sdma_v5_2_enable(adev, false);
+     sdma_v5_2_ctx_switch_disable_all(adev);
+     sdma_v5_2_halt(adev);
 
      return 0;
 }
--

2.25.1

author
#6 lore raw

Sorry about that. But can you please confirm that the call of sdma_v5_2_ctx_switch_disable_all() is still required to fully disable the engine?

Thanks,
Haohui

Quoted 297 lines · Lang Yu
Quoted 3 lines · Haohui Mai

Thanks for pointing it out. The v5 patch added the code back.

~Haohui

But you didn't remove the following codes added in v4.

        if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +3 -0
+               sdma_v5_2_ctx_switch_disable_all(adev);
+               sdma_v5_2_halt(adev);
+

sdma_v5_2_halt() has been called in sdma_v5_2_load_microcode().

No need to call these two functions here.

Regards,
Lang

Quoted 277 lines · Lang Yu
Quoted 276 lines · Haohui Mai
Quoted 9 lines

If I understand correctly, the original code will disable the HALT bit of the register mmSDMA0_F32_CNTL twice on other code paths -- one in the sdma_v5_2_ctx_switch_enable() and the other one in sdma_v5_2_enable().

The change ensures that the bit is only disabled once. Just wondering, which one is the expected behavior?

~Haohui

The HALT bit of the register mmSDMA0_F32_CNTL is not set/reset in original sdma_v5_2_ctx_switch_enable()
(https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgitlab.freedesktop.org%2Fagd5f%2Flinux%2F-%2Fblob%2Famd-staging-drm-next%2Fdrivers%2Fgpu%2Fdrm%2Famd%2Famdgpu%2Fsdma_v5_2.c%23L523&amp;data=05%7C01%7CLang.Yu%40amd.com%7C2df035ef6ad5411215a208da299c0253%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637868047861293833%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&amp;sdata=%2FLpd6TUyxSe96rEzbyvV0zXNPZ%2BIokiT8KtGAKea8Co%3D&amp;reserved=0)

Regards,
Lang

Quoted 258 lines · Lang Yu
Quoted 257 lines · On 04/28/ , [email protected] wrote:
Quoted 187 lines

From: Haohui Mai <[email protected]>

The patch fully deactivates the DMA engine before setting up the ring buffer to avoid potential data races and crashes.

The v4 patch addresses the comments from Lang Yu <[email protected]>.

---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 113 ++++++++++++++-----------
 1 file changed, 62 insertions(+), 51 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +54 -44
index 013d2dec81d0..b000117b77d0 100644
@@ -459,7 +459,6 @@ static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 se
      }
 }
 
-
 /**
  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
  *
@@ -505,17 +504,21 @@ static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
 }
 
 /**
- * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
+ * sdma_v5_2_ctx_switch_enable_for_instance - start the async dma engines
+ * context switch for an instance
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs context switch.
+ * @instance_idx: the index of the SDMA instance
  *
- * Halt or unhalt the async dma engines context switch.
+ * Unhalt the async dma engines context switch.
  */
-static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_enable_for_instance(struct amdgpu_device *adev, int instance_idx)
 {
      u32 f32_cntl, phase_quantum = 0;
-     int i;
+
+     if (WARN_ON(instance_idx >= adev->sdma.num_instances)) {
+             return;
+     }
 
      if (amdgpu_sdma_phase_quantum) {
              unsigned value = amdgpu_sdma_phase_quantum;
@@ -539,61 +542,71 @@ static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
              phase_quantum =
                      value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
                      unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
-     }
-
-     for (i = 0; i < adev->sdma.num_instances; i++) {
-             if (enable && amdgpu_sdma_phase_quantum) {
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
-                            phase_quantum);
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
-                            phase_quantum);
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
-                            phase_quantum);
-             }
 
-             if (!amdgpu_sriov_vf(adev)) {
-                     f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
-                     f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
-                                     AUTO_CTXSW_ENABLE, enable ? 1 : 0);
-                     WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
-             }
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE0_QUANTUM),
+                     phase_quantum);
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE1_QUANTUM),
+                 phase_quantum);
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE2_QUANTUM),
+                 phase_quantum);
      }
 
+     if (!amdgpu_sriov_vf(adev)) {
+             f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL));
+             f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+                             AUTO_CTXSW_ENABLE, 1);
+             WREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL), f32_cntl);
+     }
 }
 
 /**
- * sdma_v5_2_enable - stop the async dma engines
+ * sdma_v5_2_ctx_switch_disable_all - stop the async dma engines context switch
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs.
  *
- * Halt or unhalt the async dma engines.
+ * Halt the async dma engines context switch.
  */
-static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_disable_all(struct amdgpu_device *adev)
 {
      u32 f32_cntl;
      int i;
 
-     if (!enable) {
-             sdma_v5_2_gfx_stop(adev);
-             sdma_v5_2_rlc_stop(adev);
-     }
+     if (amdgpu_sriov_vf(adev))
+             return;
 
-     if (!amdgpu_sriov_vf(adev)) {
-             for (i = 0; i < adev->sdma.num_instances; i++) {
-                     f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
-                     f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
-                     WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
-             }
+     for (i = 0; i < adev->sdma.num_instances; i++) {
+             f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
+             f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+                             AUTO_CTXSW_ENABLE, 0);
+             WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
      }
 }
 
+/**
+ * sdma_v5_2_halt - stop the async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Halt the async dma engines.
+ */
+static void sdma_v5_2_halt(struct amdgpu_device *adev)
+{
+     sdma_v5_2_gfx_stop(adev);
+     sdma_v5_2_rlc_stop(adev);
+}
+
 /**
  * sdma_v5_2_gfx_resume - setup and start the async dma engines
  *
  * @adev: amdgpu_device pointer
  *
  * Set up the gfx DMA ring buffers and enable them.
+ * It assumes that the dma engine is stopped for each instance.
+ * The function enables the engine and preemptions sequentially for each instance.
+ *
  * Returns 0 for success, error for failure.
  */
 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
@@ -737,10 +750,7 @@ static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
 
              ring->sched.ready = true;
 
-             if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
-                     sdma_v5_2_ctx_switch_enable(adev, true);
-                     sdma_v5_2_enable(adev, true);
-             }
+             sdma_v5_2_ctx_switch_enable_for_instance(adev, i);
 
              r = amdgpu_ring_test_ring(ring);
              if (r) {
@@ -784,7 +794,7 @@ static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
      int i, j;
 
      /* halt the MEs */
-     sdma_v5_2_enable(adev, false);
+     sdma_v5_2_halt(adev);
 
      for (i = 0; i < adev->sdma.num_instances; i++) {
              if (!adev->sdma.instance[i].fw)
@@ -856,8 +866,8 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
      int r = 0;
 
      if (amdgpu_sriov_vf(adev)) {
-             sdma_v5_2_ctx_switch_enable(adev, false);
-             sdma_v5_2_enable(adev, false);
+             sdma_v5_2_ctx_switch_disable_all(adev);
+             sdma_v5_2_halt(adev);
 
              /* set RB registers */
              r = sdma_v5_2_gfx_resume(adev);

drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c static int sdma_v5_2_start(struct amdgpu_device *adev) · line 879

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +2 -0
@@ -865,6 +875,9 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
      }
 
      if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
+             sdma_v5_2_ctx_switch_disable_all(adev);
+             sdma_v5_2_halt(adev);

You don't need call these functions here.

In original code, you can see sdma_v5_2_load_microcode() calls sdma_v5_2_enable(adev, false) to halt the engine before load ucode.

What I mean is sdma_v5_2_halt(adev) should have no functional change with original sdma_v5_2_enable(adev, false), like this,

static void sdma_v5_2_halt(struct amdgpu_device *adev) {
        u32 f32_cntl;
        int i;

        sdma_v5_2_gfx_stop(adev);
        sdma_v5_2_rlc_stop(adev);

        if (!amdgpu_sriov_vf(adev)) {
                for (i = 0; i < adev->sdma.num_instances; i++) {
                        f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
                        f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
                        WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
                }
        }
}

But you remove the codes to set the HLAT bit of register mmSDMA0_F32_CNTL.

Regards,
Lang

Quoted 33 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +5 -8
              r = sdma_v5_2_load_microcode(adev);
              if (r)
                      return r;
@@ -881,12 +894,10 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
              amdgpu_gfx_off_ctrl(adev, false);
 
      sdma_v5_2_soft_reset(adev);
-     /* unhalt the MEs */
-     sdma_v5_2_enable(adev, true);
-     /* enable sdma ring preemption */
-     sdma_v5_2_ctx_switch_enable(adev, true);
 
-     /* start the gfx rings and rlc compute queues */
+     /* Soft reset supposes to disable the dma engine and preemption.
+      * Now start the gfx rings and rlc compute queues.
+      */
      r = sdma_v5_2_gfx_resume(adev);
      if (adev->in_s0ix)
              amdgpu_gfx_off_ctrl(adev, true);
@@ -1340,8 +1351,8 @@ static int sdma_v5_2_hw_fini(void *handle)
      if (amdgpu_sriov_vf(adev))
              return 0;
 
-     sdma_v5_2_ctx_switch_enable(adev, false);
-     sdma_v5_2_enable(adev, false);
+     sdma_v5_2_ctx_switch_disable_all(adev);
+     sdma_v5_2_halt(adev);
 
      return 0;
 }
--

2.25.1

#7 lore raw
Quoted 3 lines · Haohui Mai

Sorry about that. But can you please confirm that the call of sdma_v5_2_ctx_switch_disable_all() is still required to fully disable the engine?

For sdma_v5_2_load_microcode(), I think sdma_v5_2_halt() will suffice.

Regards,
Lang

Quoted 301 lines

Thanks,
Haohui

Quoted 297 lines · Lang Yu
Quoted 3 lines · Haohui Mai

Thanks for pointing it out. The v5 patch added the code back.

~Haohui

But you didn't remove the following codes added in v4.

        if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
(quoted diff) +3 -0
+               sdma_v5_2_ctx_switch_disable_all(adev);
+               sdma_v5_2_halt(adev);
+

sdma_v5_2_halt() has been called in sdma_v5_2_load_microcode().

No need to call these two functions here.

Regards,
Lang

Quoted 277 lines · Lang Yu
Quoted 276 lines · Haohui Mai
Quoted 9 lines

If I understand correctly, the original code will disable the HALT bit of the register mmSDMA0_F32_CNTL twice on other code paths -- one in the sdma_v5_2_ctx_switch_enable() and the other one in sdma_v5_2_enable().

The change ensures that the bit is only disabled once. Just wondering, which one is the expected behavior?

~Haohui

The HALT bit of the register mmSDMA0_F32_CNTL is not set/reset in original sdma_v5_2_ctx_switch_enable()
(https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgitlab.freedesktop.org%2Fagd5f%2Flinux%2F-%2Fblob%2Famd-staging-drm-next%2Fdrivers%2Fgpu%2Fdrm%2Famd%2Famdgpu%2Fsdma_v5_2.c%23L523&amp;data=05%7C01%7CLang.Yu%40amd.com%7Cd3ebe4ce65f844567b0708da29bf15f2%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637868197938534483%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&amp;sdata=rHBvnV49RBwwYqFp8DmaN8lNBLgw4osg2wRFlxgYKSk%3D&amp;reserved=0)

Regards,
Lang

Quoted 258 lines · Lang Yu
Quoted 257 lines · On 04/28/ , [email protected] wrote:
Quoted 187 lines

From: Haohui Mai <[email protected]>

The patch fully deactivates the DMA engine before setting up the ring buffer to avoid potential data races and crashes.

The v4 patch addresses the comments from Lang Yu <[email protected]>.

---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 113 ++++++++++++++-----------
 1 file changed, 62 insertions(+), 51 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +54 -44
index 013d2dec81d0..b000117b77d0 100644
@@ -459,7 +459,6 @@ static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 se
      }
 }
 
-
 /**
  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
  *
@@ -505,17 +504,21 @@ static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
 }
 
 /**
- * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
+ * sdma_v5_2_ctx_switch_enable_for_instance - start the async dma engines
+ * context switch for an instance
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs context switch.
+ * @instance_idx: the index of the SDMA instance
  *
- * Halt or unhalt the async dma engines context switch.
+ * Unhalt the async dma engines context switch.
  */
-static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_enable_for_instance(struct amdgpu_device *adev, int instance_idx)
 {
      u32 f32_cntl, phase_quantum = 0;
-     int i;
+
+     if (WARN_ON(instance_idx >= adev->sdma.num_instances)) {
+             return;
+     }
 
      if (amdgpu_sdma_phase_quantum) {
              unsigned value = amdgpu_sdma_phase_quantum;
@@ -539,61 +542,71 @@ static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
              phase_quantum =
                      value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
                      unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
-     }
-
-     for (i = 0; i < adev->sdma.num_instances; i++) {
-             if (enable && amdgpu_sdma_phase_quantum) {
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
-                            phase_quantum);
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
-                            phase_quantum);
-                     WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
-                            phase_quantum);
-             }
 
-             if (!amdgpu_sriov_vf(adev)) {
-                     f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
-                     f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
-                                     AUTO_CTXSW_ENABLE, enable ? 1 : 0);
-                     WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
-             }
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE0_QUANTUM),
+                     phase_quantum);
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE1_QUANTUM),
+                 phase_quantum);
+             WREG32_SOC15_IP(GC,
+                     sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE2_QUANTUM),
+                 phase_quantum);
      }
 
+     if (!amdgpu_sriov_vf(adev)) {
+             f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL));
+             f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+                             AUTO_CTXSW_ENABLE, 1);
+             WREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL), f32_cntl);
+     }
 }
 
 /**
- * sdma_v5_2_enable - stop the async dma engines
+ * sdma_v5_2_ctx_switch_disable_all - stop the async dma engines context switch
  *
  * @adev: amdgpu_device pointer
- * @enable: enable/disable the DMA MEs.
  *
- * Halt or unhalt the async dma engines.
+ * Halt the async dma engines context switch.
  */
-static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
+static void sdma_v5_2_ctx_switch_disable_all(struct amdgpu_device *adev)
 {
      u32 f32_cntl;
      int i;
 
-     if (!enable) {
-             sdma_v5_2_gfx_stop(adev);
-             sdma_v5_2_rlc_stop(adev);
-     }
+     if (amdgpu_sriov_vf(adev))
+             return;
 
-     if (!amdgpu_sriov_vf(adev)) {
-             for (i = 0; i < adev->sdma.num_instances; i++) {
-                     f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
-                     f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
-                     WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
-             }
+     for (i = 0; i < adev->sdma.num_instances; i++) {
+             f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
+             f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+                             AUTO_CTXSW_ENABLE, 0);
+             WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
      }
 }
 
+/**
+ * sdma_v5_2_halt - stop the async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Halt the async dma engines.
+ */
+static void sdma_v5_2_halt(struct amdgpu_device *adev)
+{
+     sdma_v5_2_gfx_stop(adev);
+     sdma_v5_2_rlc_stop(adev);
+}
+
 /**
  * sdma_v5_2_gfx_resume - setup and start the async dma engines
  *
  * @adev: amdgpu_device pointer
  *
  * Set up the gfx DMA ring buffers and enable them.
+ * It assumes that the dma engine is stopped for each instance.
+ * The function enables the engine and preemptions sequentially for each instance.
+ *
  * Returns 0 for success, error for failure.
  */
 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
@@ -737,10 +750,7 @@ static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
 
              ring->sched.ready = true;
 
-             if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
-                     sdma_v5_2_ctx_switch_enable(adev, true);
-                     sdma_v5_2_enable(adev, true);
-             }
+             sdma_v5_2_ctx_switch_enable_for_instance(adev, i);
 
              r = amdgpu_ring_test_ring(ring);
              if (r) {
@@ -784,7 +794,7 @@ static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
      int i, j;
 
      /* halt the MEs */
-     sdma_v5_2_enable(adev, false);
+     sdma_v5_2_halt(adev);
 
      for (i = 0; i < adev->sdma.num_instances; i++) {
              if (!adev->sdma.instance[i].fw)
@@ -856,8 +866,8 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
      int r = 0;
 
      if (amdgpu_sriov_vf(adev)) {
-             sdma_v5_2_ctx_switch_enable(adev, false);
-             sdma_v5_2_enable(adev, false);
+             sdma_v5_2_ctx_switch_disable_all(adev);
+             sdma_v5_2_halt(adev);
 
              /* set RB registers */
              r = sdma_v5_2_gfx_resume(adev);

drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c static int sdma_v5_2_start(struct amdgpu_device *adev) · line 879

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +2 -0
@@ -865,6 +875,9 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
      }
 
      if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
+             sdma_v5_2_ctx_switch_disable_all(adev);
+             sdma_v5_2_halt(adev);

You don't need call these functions here.

In original code, you can see sdma_v5_2_load_microcode() calls sdma_v5_2_enable(adev, false) to halt the engine before load ucode.

What I mean is sdma_v5_2_halt(adev) should have no functional change with original sdma_v5_2_enable(adev, false), like this,

static void sdma_v5_2_halt(struct amdgpu_device *adev) {
        u32 f32_cntl;
        int i;

        sdma_v5_2_gfx_stop(adev);
        sdma_v5_2_rlc_stop(adev);

        if (!amdgpu_sriov_vf(adev)) {
                for (i = 0; i < adev->sdma.num_instances; i++) {
                        f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
                        f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
                        WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
                }
        }
}

But you remove the codes to set the HLAT bit of register mmSDMA0_F32_CNTL.

Regards,
Lang

Quoted 33 lines
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +5 -8
              r = sdma_v5_2_load_microcode(adev);
              if (r)
                      return r;
@@ -881,12 +894,10 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
              amdgpu_gfx_off_ctrl(adev, false);
 
      sdma_v5_2_soft_reset(adev);
-     /* unhalt the MEs */
-     sdma_v5_2_enable(adev, true);
-     /* enable sdma ring preemption */
-     sdma_v5_2_ctx_switch_enable(adev, true);
 
-     /* start the gfx rings and rlc compute queues */
+     /* Soft reset supposes to disable the dma engine and preemption.
+      * Now start the gfx rings and rlc compute queues.
+      */
      r = sdma_v5_2_gfx_resume(adev);
      if (adev->in_s0ix)
              amdgpu_gfx_off_ctrl(adev, true);
@@ -1340,8 +1351,8 @@ static int sdma_v5_2_hw_fini(void *handle)
      if (amdgpu_sriov_vf(adev))
              return 0;
 
-     sdma_v5_2_ctx_switch_enable(adev, false);
-     sdma_v5_2_enable(adev, false);
+     sdma_v5_2_ctx_switch_disable_all(adev);
+     sdma_v5_2_halt(adev);
 
      return 0;
 }
--

2.25.1