Skip to contentSkip to conversation

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

PATCH In review [email protected] 2 participants 11 messages Apr 30, 2022 – May 11, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v6 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.

---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 109 +++++++++++++++----------
 1 file changed, 64 insertions(+), 45 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +64 -45
index 013d2dec81d0..1fac9d8e99de 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;
@@ -539,50 +542,68 @@ 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;
+
+	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)
+{
+	int i;
+	u32 f32_cntl;
+
+	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, enable ? 0 : 1);
+			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);
 		}
 	}
@@ -594,6 +615,9 @@ static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
  * @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 +761,7 @@ 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 +805,7 @@ 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 +877,8 @@ 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);
@@ -881,12 +902,10 @@ 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 +1359,8 @@ 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
#1 lore raw

Ping...
________________________________
From: [email protected] <[email protected]>
Sent: Saturday, April 30, 2022 3:34:00 PM
To: [email protected] <[email protected]>
Cc: [email protected] <[email protected]>; [email protected] <[email protected]>; [email protected] <[email protected]>; [email protected] <[email protected]>; [email protected] <[email protected]>; Haohui Mai <[email protected]>
Subject: [PATCH v6] 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.

---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 109 +++++++++++++++----------
 1 file changed, 64 insertions(+), 45 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +64 -45
index 013d2dec81d0..1fac9d8e99de 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;
@@ -539,50 +542,68 @@ 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;
+
+       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)
+{
+       int i;
+       u32 f32_cntl;
+
+       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, enable ? 0 : 1);
+                       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);
                 }
         }
@@ -594,6 +615,9 @@ static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
  * @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 +761,7 @@ 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 +805,7 @@ 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 +877,8 @@ 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);
@@ -881,12 +902,10 @@ 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 +1359,8 @@ 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

#2 lore raw
In reply to [email protected]
Quoted 5 lines · On Sat, Apr 30, 2022 at 3:34 AM <[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.

Does this actually fix an issue you are seeing? I don't think it will hurt anything, but I also don't think it's strictly necessary. AFAIK, only the HALT bit really matters. So the commit message might be somewhat misleading.

Alex

Quoted 211 lines
---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 109 +++++++++++++++----------
 1 file changed, 64 insertions(+), 45 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +64 -46
index 013d2dec81d0..1fac9d8e99de 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;
@@ -539,50 +542,68 @@ 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;
+
+       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)
+{
+       int i;
+       u32 f32_cntl;
+
+       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, enable ? 0 : 1);
+                       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);
                }
        }
@@ -594,6 +615,9 @@ static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
  * @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 +761,7 @@ 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 +805,7 @@ 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 +877,8 @@ 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);
@@ -881,12 +902,10 @@ 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 +1359,8 @@ 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
#3 lore raw

The only thing that matters is that the IP should be halted before programming the ring buffers.

What about rephrasing the commit messages to highlight the issue a little bit better?

Quoted 226 lines · Alex Deucher
Quoted 5 lines · On Sat, Apr 30, 2022 at 3:34 AM <[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.

Does this actually fix an issue you are seeing? I don't think it will hurt anything, but I also don't think it's strictly necessary. AFAIK, only the HALT bit really matters. So the commit message might be somewhat misleading.

Alex

Quoted 211 lines
---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 109 +++++++++++++++----------
 1 file changed, 64 insertions(+), 45 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +64 -46
index 013d2dec81d0..1fac9d8e99de 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;
@@ -539,50 +542,68 @@ 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;
+
+       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)
+{
+       int i;
+       u32 f32_cntl;
+
+       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, enable ? 0 : 1);
+                       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);
                }
        }
@@ -594,6 +615,9 @@ static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
  * @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 +761,7 @@ 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 +805,7 @@ 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 +877,8 @@ 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);
@@ -881,12 +902,10 @@ 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 +1359,8 @@ 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

#4 lore raw
Quoted 6 lines · Haohui Mai

The only thing that matters is that the IP should be halted before programming the ring buffers.

What about rephrasing the commit messages to highlight the issue a little bit better?

Yeah, that is fine. Thanks!

Alex

Quoted 228 lines · Alex Deucher
Quoted 226 lines · On Sat, Apr 30, 2022 at 3:34 AM <[email protected]> wrote:
Quoted 5 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.

Does this actually fix an issue you are seeing? I don't think it will hurt anything, but I also don't think it's strictly necessary. AFAIK, only the HALT bit really matters. So the commit message might be somewhat misleading.

Alex

Quoted 211 lines
---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 109 +++++++++++++++----------
 1 file changed, 64 insertions(+), 45 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +64 -46
index 013d2dec81d0..1fac9d8e99de 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;
@@ -539,50 +542,68 @@ 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;
+
+       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)
+{
+       int i;
+       u32 f32_cntl;
+
+       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, enable ? 0 : 1);
+                       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);
                }
        }
@@ -594,6 +615,9 @@ static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
  * @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 +761,7 @@ 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 +805,7 @@ 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 +877,8 @@ 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);
@@ -881,12 +902,10 @@ 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 +1359,8 @@ 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
#5 lore raw

What about

Setting the HALT bit of SDMA_F32_CNTL in all paths before programming the ring buffer of the SDMA engine.

No other changes are required in the patch.

~Haohui

Quoted 241 lines · Alex Deucher
Quoted 6 lines · Haohui Mai

The only thing that matters is that the IP should be halted before programming the ring buffers.

What about rephrasing the commit messages to highlight the issue a little bit better?

Yeah, that is fine. Thanks!

Alex

Quoted 228 lines · Alex Deucher
Quoted 226 lines · On Sat, Apr 30, 2022 at 3:34 AM <[email protected]> wrote:
Quoted 5 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.

Does this actually fix an issue you are seeing? I don't think it will hurt anything, but I also don't think it's strictly necessary. AFAIK, only the HALT bit really matters. So the commit message might be somewhat misleading.

Alex

Quoted 211 lines
---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 109 +++++++++++++++----------
 1 file changed, 64 insertions(+), 45 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +64 -46
index 013d2dec81d0..1fac9d8e99de 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;
@@ -539,50 +542,68 @@ 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;
+
+       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)
+{
+       int i;
+       u32 f32_cntl;
+
+       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, enable ? 0 : 1);
+                       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);
                }
        }
@@ -594,6 +615,9 @@ static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
  * @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 +761,7 @@ 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 +805,7 @@ 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 +877,8 @@ 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);
@@ -881,12 +902,10 @@ 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 +1359,8 @@ 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

#6 lore raw
Quoted 6 lines · Haohui Mai

What about

Setting the HALT bit of SDMA_F32_CNTL in all paths before programming the ring buffer of the SDMA engine.

Sounds fine to me.

Alex

Quoted 246 lines

No other changes are required in the patch.

~Haohui

Quoted 241 lines · Alex Deucher
Quoted 6 lines · Haohui Mai

The only thing that matters is that the IP should be halted before programming the ring buffers.

What about rephrasing the commit messages to highlight the issue a little bit better?

Yeah, that is fine. Thanks!

Alex

Quoted 228 lines · Alex Deucher
Quoted 226 lines · On Sat, Apr 30, 2022 at 3:34 AM <[email protected]> wrote:
Quoted 5 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.

Does this actually fix an issue you are seeing? I don't think it will hurt anything, but I also don't think it's strictly necessary. AFAIK, only the HALT bit really matters. So the commit message might be somewhat misleading.

Alex

Quoted 211 lines
---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 109 +++++++++++++++----------
 1 file changed, 64 insertions(+), 45 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +64 -46
index 013d2dec81d0..1fac9d8e99de 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;
@@ -539,50 +542,68 @@ 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;
+
+       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)
+{
+       int i;
+       u32 f32_cntl;
+
+       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, enable ? 0 : 1);
+                       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);
                }
        }
@@ -594,6 +615,9 @@ static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
  * @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 +761,7 @@ 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 +805,7 @@ 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 +877,8 @@ 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);
@@ -881,12 +902,10 @@ 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 +1359,8 @@ 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
#7 lore raw

Hi Alex,

Is there anything open before it can be merged?

Thanks,
Haohui

Quoted 259 lines · Alex Deucher
Quoted 6 lines · Haohui Mai

What about

Setting the HALT bit of SDMA_F32_CNTL in all paths before programming the ring buffer of the SDMA engine.

Sounds fine to me.

Alex

Quoted 246 lines

No other changes are required in the patch.

~Haohui

Quoted 241 lines · Alex Deucher
Quoted 6 lines · Haohui Mai

The only thing that matters is that the IP should be halted before programming the ring buffers.

What about rephrasing the commit messages to highlight the issue a little bit better?

Yeah, that is fine. Thanks!

Alex

Quoted 228 lines · Alex Deucher
Quoted 226 lines · On Sat, Apr 30, 2022 at 3:34 AM <[email protected]> wrote:
Quoted 5 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.

Does this actually fix an issue you are seeing? I don't think it will hurt anything, but I also don't think it's strictly necessary. AFAIK, only the HALT bit really matters. So the commit message might be somewhat misleading.

Alex

Quoted 211 lines
---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 109 +++++++++++++++----------
 1 file changed, 64 insertions(+), 45 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +64 -46
index 013d2dec81d0..1fac9d8e99de 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;
@@ -539,50 +542,68 @@ 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;
+
+       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)
+{
+       int i;
+       u32 f32_cntl;
+
+       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, enable ? 0 : 1);
+                       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);
                }
        }
@@ -594,6 +615,9 @@ static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
  * @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 +761,7 @@ 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 +805,7 @@ 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 +877,8 @@ 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);
@@ -881,12 +902,10 @@ 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 +1359,8 @@ 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

#8 lore raw
Quoted 4 lines · Haohui Mai

Hi Alex,

Is there anything open before it can be merged?

Were you going to send an updated patch?

Alex

Quoted 264 lines

Thanks,
Haohui

Quoted 259 lines · Alex Deucher
Quoted 6 lines · Haohui Mai

What about

Setting the HALT bit of SDMA_F32_CNTL in all paths before programming the ring buffer of the SDMA engine.

Sounds fine to me.

Alex

Quoted 246 lines

No other changes are required in the patch.

~Haohui

Quoted 241 lines · Alex Deucher
Quoted 6 lines · Haohui Mai

The only thing that matters is that the IP should be halted before programming the ring buffers.

What about rephrasing the commit messages to highlight the issue a little bit better?

Yeah, that is fine. Thanks!

Alex

Quoted 228 lines · Alex Deucher
Quoted 226 lines · On Sat, Apr 30, 2022 at 3:34 AM <[email protected]> wrote:
Quoted 5 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.

Does this actually fix an issue you are seeing? I don't think it will hurt anything, but I also don't think it's strictly necessary. AFAIK, only the HALT bit really matters. So the commit message might be somewhat misleading.

Alex

Quoted 211 lines
---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 109 +++++++++++++++----------
 1 file changed, 64 insertions(+), 45 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +64 -46
index 013d2dec81d0..1fac9d8e99de 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;
@@ -539,50 +542,68 @@ 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;
+
+       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)
+{
+       int i;
+       u32 f32_cntl;
+
+       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, enable ? 0 : 1);
+                       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);
                }
        }
@@ -594,6 +615,9 @@ static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
  * @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 +761,7 @@ 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 +805,7 @@ 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 +877,8 @@ 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);
@@ -881,12 +902,10 @@ 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 +1359,8 @@ 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
#9 lore raw

It should be an identical patch except for the commit message. Do you want me to send out a new one? Either way is fine with me.

~Haohui

Quoted 275 lines · Alex Deucher
Quoted 4 lines · Haohui Mai

Hi Alex,

Is there anything open before it can be merged?

Were you going to send an updated patch?

Alex

Quoted 264 lines

Thanks,
Haohui

Quoted 259 lines · Alex Deucher
Quoted 6 lines · Haohui Mai

What about

Setting the HALT bit of SDMA_F32_CNTL in all paths before programming the ring buffer of the SDMA engine.

Sounds fine to me.

Alex

Quoted 246 lines

No other changes are required in the patch.

~Haohui

Quoted 241 lines · Alex Deucher
Quoted 6 lines · Haohui Mai

The only thing that matters is that the IP should be halted before programming the ring buffers.

What about rephrasing the commit messages to highlight the issue a little bit better?

Yeah, that is fine. Thanks!

Alex

Quoted 228 lines · Alex Deucher
Quoted 226 lines · On Sat, Apr 30, 2022 at 3:34 AM <[email protected]> wrote:
Quoted 5 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.

Does this actually fix an issue you are seeing? I don't think it will hurt anything, but I also don't think it's strictly necessary. AFAIK, only the HALT bit really matters. So the commit message might be somewhat misleading.

Alex

Quoted 211 lines
---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 109 +++++++++++++++----------
 1 file changed, 64 insertions(+), 45 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +64 -46
index 013d2dec81d0..1fac9d8e99de 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;
@@ -539,50 +542,68 @@ 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;
+
+       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)
+{
+       int i;
+       u32 f32_cntl;
+
+       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, enable ? 0 : 1);
+                       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);
                }
        }
@@ -594,6 +615,9 @@ static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
  * @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 +761,7 @@ 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 +805,7 @@ 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 +877,8 @@ 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);
@@ -881,12 +902,10 @@ 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 +1359,8 @@ 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

#10 lore raw

I've applied it with the change to the commit message. Sorry for the confusion.

Alex

Quoted 282 lines · Haohui Mai

It should be an identical patch except for the commit message. Do you want me to send out a new one? Either way is fine with me.

~Haohui

Quoted 275 lines · Alex Deucher
Quoted 4 lines · Haohui Mai

Hi Alex,

Is there anything open before it can be merged?

Were you going to send an updated patch?

Alex

Quoted 264 lines

Thanks,
Haohui

Quoted 259 lines · Alex Deucher
Quoted 6 lines · Haohui Mai

What about

Setting the HALT bit of SDMA_F32_CNTL in all paths before programming the ring buffer of the SDMA engine.

Sounds fine to me.

Alex

Quoted 246 lines

No other changes are required in the patch.

~Haohui

Quoted 241 lines · Alex Deucher
Quoted 6 lines · Haohui Mai

The only thing that matters is that the IP should be halted before programming the ring buffers.

What about rephrasing the commit messages to highlight the issue a little bit better?

Yeah, that is fine. Thanks!

Alex

Quoted 228 lines · Alex Deucher
Quoted 226 lines · On Sat, Apr 30, 2022 at 3:34 AM <[email protected]> wrote:
Quoted 5 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.

Does this actually fix an issue you are seeing? I don't think it will hurt anything, but I also don't think it's strictly necessary. AFAIK, only the HALT bit really matters. So the commit message might be somewhat misleading.

Alex

Quoted 211 lines
---
 drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 109 +++++++++++++++----------
 1 file changed, 64 insertions(+), 45 deletions(-)
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c +64 -46
index 013d2dec81d0..1fac9d8e99de 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;
@@ -539,50 +542,68 @@ 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;
+
+       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)
+{
+       int i;
+       u32 f32_cntl;
+
+       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, enable ? 0 : 1);
+                       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);
                }
        }
@@ -594,6 +615,9 @@ static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
  * @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 +761,7 @@ 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 +805,7 @@ 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 +877,8 @@ 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);
@@ -881,12 +902,10 @@ 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 +1359,8 @@ 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