Skip to contentSkip to conversation

drm/amdgpu: Use driver mode reset for data poison

PATCH In review Hawking Zhang 3 participants 4 messages Apr 16, 2024

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/amdgpu: Use driver mode reset for data poison

mode-2 reset is the only reliable method that can get GC/SDMA back when poison is consumed. mmhub requires mode-1 reset.

---
 .../gpu/drm/amd/amdkfd/kfd_int_process_v9.c   | 27 ++++++-------------
 1 file changed, 8 insertions(+), 19 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +8 -19
index c368c70df3f4a..c3beb872adf8d 100644
@@ -144,7 +144,7 @@ static void event_interrupt_poison_consumption_v9(struct kfd_node *dev,
 				uint16_t pasid, uint16_t client_id)
 {
 	enum amdgpu_ras_block block = 0;
-	int old_poison, ret = -EINVAL;
+	int old_poison;
 	uint32_t reset = 0;
 	struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
 
@@ -163,17 +163,13 @@ static void event_interrupt_poison_consumption_v9(struct kfd_node *dev,
 	case SOC15_IH_CLIENTID_SE2SH:
 	case SOC15_IH_CLIENTID_SE3SH:
 	case SOC15_IH_CLIENTID_UTCL2:
-		ret = kfd_dqm_evict_pasid(dev->dqm, pasid);
 		block = AMDGPU_RAS_BLOCK__GFX;
-		if (ret)
-			reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
+		reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
 		break;
 	case SOC15_IH_CLIENTID_VMC:
 	case SOC15_IH_CLIENTID_VMC1:
-		ret = kfd_dqm_evict_pasid(dev->dqm, pasid);
 		block = AMDGPU_RAS_BLOCK__MMHUB;
-		if (ret)
-			reset = AMDGPU_RAS_GPU_RESET_MODE1_RESET;
+		reset = AMDGPU_RAS_GPU_RESET_MODE1_RESET;
 		break;
 	case SOC15_IH_CLIENTID_SDMA0:
 	case SOC15_IH_CLIENTID_SDMA1:
@@ -184,22 +180,15 @@ static void event_interrupt_poison_consumption_v9(struct kfd_node *dev,
 		reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
 		break;
 	default:
-		break;
+		dev_warn(dev->adev->dev,
+			 "client %d does not support poison consumption\n", client_id);
+		return;
 	}
 
 	kfd_signal_poison_consumed_event(dev, pasid);
 
-	/* resetting queue passes, do page retirement without gpu reset
-	 * resetting queue fails, fallback to gpu reset solution
-	 */
-	if (!ret)
-		dev_warn(dev->adev->dev,
-			"RAS poison consumption, unmap queue flow succeeded: client id %d\n",
-			client_id);
-	else
-		dev_warn(dev->adev->dev,
-			"RAS poison consumption, fall back to gpu reset flow: client id %d\n",
-			client_id);
+	dev_warn(dev->adev->dev,
+		 "poison is consumed by client %d, kick off gpu reset flow\n", client_id);
 
 	amdgpu_amdkfd_ras_poison_consumption_handler(dev->adev, block, reset);
 }
Signature · 1 line
2.17.1
Reviewed-by on v1 1/1
#1 lore raw

[AMD Official Use Only - General]

Quoted 86 lines

-----Original Message-----
From: Hawking Zhang <[email protected]>
Sent: Tuesday, April 16, 2024 2:16 PM
To: [email protected]; Zhou1, Tao <[email protected]>
Cc: Zhang, Hawking <[email protected]>
Subject: [PATCH] drm/amdgpu: Use driver mode reset for data poison

mode-2 reset is the only reliable method that can get GC/SDMA back when poison is consumed. mmhub requires
mode-1 reset.

---
 .../gpu/drm/amd/amdkfd/kfd_int_process_v9.c   | 27 ++++++-------------
 1 file changed, 8 insertions(+), 19 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +3 -7
index c368c70df3f4a..c3beb872adf8d 100644
@@ -144,7 +144,7 @@ static void
 event_interrupt_poison_consumption_v9(struct kfd_node *dev,
                              uint16_t pasid, uint16_t client_id)  {
      enum amdgpu_ras_block block = 0;
-     int old_poison, ret = -EINVAL;
+     int old_poison;
      uint32_t reset = 0;
      struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
 
@@ -163,17 +163,13 @@ static void
 event_interrupt_poison_consumption_v9(struct kfd_node *dev,
      case SOC15_IH_CLIENTID_SE2SH:
      case SOC15_IH_CLIENTID_SE3SH:
      case SOC15_IH_CLIENTID_UTCL2:
-             ret = kfd_dqm_evict_pasid(dev->dqm, pasid);
              block = AMDGPU_RAS_BLOCK__GFX;
-             if (ret)
-                     reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
+             reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
              break;
      case SOC15_IH_CLIENTID_VMC:
      case SOC15_IH_CLIENTID_VMC1:
-             ret = kfd_dqm_evict_pasid(dev->dqm, pasid);
              block = AMDGPU_RAS_BLOCK__MMHUB;
-             if (ret)
-                     reset = AMDGPU_RAS_GPU_RESET_MODE1_RESET;
+             reset = AMDGPU_RAS_GPU_RESET_MODE1_RESET;
              break;
      case SOC15_IH_CLIENTID_SDMA0:

case SOC15_IH_CLIENTID_SDMA1:

drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +6 -12
@@ -184,22 +180,15 @@ static void
 event_interrupt_poison_consumption_v9(struct kfd_node *dev,
              reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
              break;
      default:
-             break;
+             dev_warn(dev->adev->dev,
+                      "client %d does not support poison consumption\n",
 client_id);
+             return;
      }
 
      kfd_signal_poison_consumed_event(dev, pasid);
 
-     /* resetting queue passes, do page retirement without gpu reset
-      * resetting queue fails, fallback to gpu reset solution
-      */
-     if (!ret)
-             dev_warn(dev->adev->dev,
-                     "RAS poison consumption, unmap queue flow
 succeeded: client id %d\n",
-                     client_id);
-     else
-             dev_warn(dev->adev->dev,
-                     "RAS poison consumption, fall back to gpu reset flow:
 client id %d\n",
-                     client_id);
+     dev_warn(dev->adev->dev,
+              "poison is consumed by client %d, kick off gpu reset flow\n",
+client_id);
 
      amdgpu_amdkfd_ras_poison_consumption_handler(dev->adev, block,

reset); }
--
2.17.1

#2 lore raw
In reply to Hawking Zhang

[Public]

Is this applicable for aldebaran also?

Thanks,
Lijo
-----Original Message-----
From: amd-gfx <[email protected]> On Behalf Of Hawking Zhang
Sent: Tuesday, April 16, 2024 11:46 AM
To: [email protected]; Zhou1, Tao <[email protected]>
Cc: Zhang, Hawking <[email protected]>
Subject: [PATCH] drm/amdgpu: Use driver mode reset for data poison

mode-2 reset is the only reliable method that can get GC/SDMA back when poison is consumed. mmhub requires mode-1 reset.

---
 .../gpu/drm/amd/amdkfd/kfd_int_process_v9.c   | 27 ++++++-------------
 1 file changed, 8 insertions(+), 19 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +9 -20
index c368c70df3f4a..c3beb872adf8d 100644
@@ -144,7 +144,7 @@ static void event_interrupt_poison_consumption_v9(struct kfd_node *dev,
                                uint16_t pasid, uint16_t client_id)  {
        enum amdgpu_ras_block block = 0;
-       int old_poison, ret = -EINVAL;
+       int old_poison;
        uint32_t reset = 0;
        struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
 
@@ -163,17 +163,13 @@ static void event_interrupt_poison_consumption_v9(struct kfd_node *dev,
        case SOC15_IH_CLIENTID_SE2SH:
        case SOC15_IH_CLIENTID_SE3SH:
        case SOC15_IH_CLIENTID_UTCL2:
-               ret = kfd_dqm_evict_pasid(dev->dqm, pasid);
                block = AMDGPU_RAS_BLOCK__GFX;
-               if (ret)
-                       reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
+               reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
                break;
        case SOC15_IH_CLIENTID_VMC:
        case SOC15_IH_CLIENTID_VMC1:
-               ret = kfd_dqm_evict_pasid(dev->dqm, pasid);
                block = AMDGPU_RAS_BLOCK__MMHUB;
-               if (ret)
-                       reset = AMDGPU_RAS_GPU_RESET_MODE1_RESET;
+               reset = AMDGPU_RAS_GPU_RESET_MODE1_RESET;
                break;
        case SOC15_IH_CLIENTID_SDMA0:
        case SOC15_IH_CLIENTID_SDMA1:
@@ -184,22 +180,15 @@ static void event_interrupt_poison_consumption_v9(struct kfd_node *dev,
                reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
                break;
        default:
-               break;
+               dev_warn(dev->adev->dev,
+                        "client %d does not support poison consumption\n", client_id);
+               return;
        }
 
        kfd_signal_poison_consumed_event(dev, pasid);
 
-       /* resetting queue passes, do page retirement without gpu reset
-        * resetting queue fails, fallback to gpu reset solution
-        */
-       if (!ret)
-               dev_warn(dev->adev->dev,
-                       "RAS poison consumption, unmap queue flow succeeded: client id %d\n",
-                       client_id);
-       else
-               dev_warn(dev->adev->dev,
-                       "RAS poison consumption, fall back to gpu reset flow: client id %d\n",
-                       client_id);
+       dev_warn(dev->adev->dev,
+                "poison is consumed by client %d, kick off gpu reset flow\n",
+client_id);
 
        amdgpu_amdkfd_ras_poison_consumption_handler(dev->adev, block, reset);  }
--

2.17.1

#3 lore raw

[Public]

Yes, the same strategy

-----Original Message-----
From: Lazar, Lijo <[email protected]>
Sent: Tuesday, April 16, 2024 02:24
To: Zhang, Hawking <[email protected]>; [email protected]; Zhou1, Tao <[email protected]>
Cc: Zhang, Hawking <[email protected]>
Subject: RE: [PATCH] drm/amdgpu: Use driver mode reset for data poison

[Public]

Is this applicable for aldebaran also?

Thanks,
Lijo
-----Original Message-----
From: amd-gfx <[email protected]> On Behalf Of Hawking Zhang
Sent: Tuesday, April 16, 2024 11:46 AM
To: [email protected]; Zhou1, Tao <[email protected]>
Cc: Zhang, Hawking <[email protected]>
Subject: [PATCH] drm/amdgpu: Use driver mode reset for data poison

mode-2 reset is the only reliable method that can get GC/SDMA back when poison is consumed. mmhub requires mode-1 reset.

---
 .../gpu/drm/amd/amdkfd/kfd_int_process_v9.c   | 27 ++++++-------------
 1 file changed, 8 insertions(+), 19 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +9 -20
index c368c70df3f4a..c3beb872adf8d 100644
@@ -144,7 +144,7 @@ static void event_interrupt_poison_consumption_v9(struct kfd_node *dev,
                                uint16_t pasid, uint16_t client_id)  {
        enum amdgpu_ras_block block = 0;
-       int old_poison, ret = -EINVAL;
+       int old_poison;
        uint32_t reset = 0;
        struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
 
@@ -163,17 +163,13 @@ static void event_interrupt_poison_consumption_v9(struct kfd_node *dev,
        case SOC15_IH_CLIENTID_SE2SH:
        case SOC15_IH_CLIENTID_SE3SH:
        case SOC15_IH_CLIENTID_UTCL2:
-               ret = kfd_dqm_evict_pasid(dev->dqm, pasid);
                block = AMDGPU_RAS_BLOCK__GFX;
-               if (ret)
-                       reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
+               reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
                break;
        case SOC15_IH_CLIENTID_VMC:
        case SOC15_IH_CLIENTID_VMC1:
-               ret = kfd_dqm_evict_pasid(dev->dqm, pasid);
                block = AMDGPU_RAS_BLOCK__MMHUB;
-               if (ret)
-                       reset = AMDGPU_RAS_GPU_RESET_MODE1_RESET;
+               reset = AMDGPU_RAS_GPU_RESET_MODE1_RESET;
                break;
        case SOC15_IH_CLIENTID_SDMA0:
        case SOC15_IH_CLIENTID_SDMA1:
@@ -184,22 +180,15 @@ static void event_interrupt_poison_consumption_v9(struct kfd_node *dev,
                reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
                break;
        default:
-               break;
+               dev_warn(dev->adev->dev,
+                        "client %d does not support poison consumption\n", client_id);
+               return;
        }
 
        kfd_signal_poison_consumed_event(dev, pasid);
 
-       /* resetting queue passes, do page retirement without gpu reset
-        * resetting queue fails, fallback to gpu reset solution
-        */
-       if (!ret)
-               dev_warn(dev->adev->dev,
-                       "RAS poison consumption, unmap queue flow succeeded: client id %d\n",
-                       client_id);
-       else
-               dev_warn(dev->adev->dev,
-                       "RAS poison consumption, fall back to gpu reset flow: client id %d\n",
-                       client_id);
+       dev_warn(dev->adev->dev,
+                "poison is consumed by client %d, kick off gpu reset
+flow\n", client_id);
 
        amdgpu_amdkfd_ras_poison_consumption_handler(dev->adev, block, reset);  }
--

2.17.1