drm/amdgpu: use scheduler fault instead of reset work

Signal a fault to the scheduler on an illegal instruction or register
access violation instead of kicking of the reset handler directly.
Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

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

@@ -830,7 +830,6 @@ struct amdgpu_device {
 	bool				need_dma32;
 	bool				need_swiotlb;
 	bool				accel_working;
-	struct work_struct		reset_work;
 	struct notifier_block		acpi_nb;
 	struct amdgpu_i2c_chan		*i2c_bus[AMDGPU_MAX_I2C_BUS];
 	struct amdgpu_debugfs		debugfs[AMDGPU_DEBUGFS_MAX_COMPONENTS];

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

@@ -93,23 +93,6 @@ static void amdgpu_hotplug_work_func(struct work_struct *work)
 	drm_helper_hpd_irq_event(dev);
 }
 
-/**
- * amdgpu_irq_reset_work_func - execute GPU reset
- *
- * @work: work struct pointer
- *
- * Execute scheduled GPU reset (Cayman+).
- * This function is called when the IRQ handler thinks we need a GPU reset.
- */
-static void amdgpu_irq_reset_work_func(struct work_struct *work)
-{
-	struct amdgpu_device *adev = container_of(work, struct amdgpu_device,
-						  reset_work);
-
-	if (!amdgpu_sriov_vf(adev) && amdgpu_device_should_recover_gpu(adev))
-		amdgpu_device_gpu_recover(adev, NULL);
-}
-
 /**
  * amdgpu_irq_disable_all - disable *all* interrupts
  *
@@ -262,15 +245,12 @@ int amdgpu_irq_init(struct amdgpu_device *adev)
 				amdgpu_hotplug_work_func);
 	}
 
-	INIT_WORK(&adev->reset_work, amdgpu_irq_reset_work_func);
-
 	adev->irq.installed = true;
 	r = drm_irq_install(adev->ddev, adev->ddev->pdev->irq);
 	if (r) {
 		adev->irq.installed = false;
 		if (!amdgpu_device_has_dc_support(adev))
 			flush_work(&adev->hotplug_work);
-		cancel_work_sync(&adev->reset_work);
 		return r;
 	}
 	adev->ddev->max_vblank_count = 0x00ffffff;
@@ -299,7 +279,6 @@ void amdgpu_irq_fini(struct amdgpu_device *adev)
 			pci_disable_msi(adev->pdev);
 		if (!amdgpu_device_has_dc_support(adev))
 			flush_work(&adev->hotplug_work);
-		cancel_work_sync(&adev->reset_work);
 	}
 
 	for (i = 0; i < AMDGPU_IRQ_CLIENTID_MAX; ++i) {

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

@@ -1214,8 +1214,11 @@ static int cik_sdma_process_illegal_inst_irq(struct amdgpu_device *adev,
 					     struct amdgpu_irq_src *source,
 					     struct amdgpu_iv_entry *entry)
 {
+	u8 instance_id;
+
 	DRM_ERROR("Illegal instruction in SDMA command stream\n");
-	schedule_work(&adev->reset_work);
+	instance_id = (entry->ring_id & 0x3) >> 0;
+	drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
 	return 0;
 }
 

drivers/gpu/drm/amd/amdgpu/gfx_v6_0.c

@@ -3393,12 +3393,31 @@ static int gfx_v6_0_eop_irq(struct amdgpu_device *adev,
 	return 0;
 }
 
+static void gfx_v6_0_fault(struct amdgpu_device *adev,
+			   struct amdgpu_iv_entry *entry)
+{
+	struct amdgpu_ring *ring;
+
+	switch (entry->ring_id) {
+	case 0:
+		ring = &adev->gfx.gfx_ring[0];
+		break;
+	case 1:
+	case 2:
+		ring = &adev->gfx.compute_ring[entry->ring_id - 1];
+		break;
+	default:
+		return;
+	}
+	drm_sched_fault(&ring->sched);
+}
+
 static int gfx_v6_0_priv_reg_irq(struct amdgpu_device *adev,
 				 struct amdgpu_irq_src *source,
 				 struct amdgpu_iv_entry *entry)
 {
 	DRM_ERROR("Illegal register access in command stream\n");
-	schedule_work(&adev->reset_work);
+	gfx_v6_0_fault(adev, entry);
 	return 0;
 }
 
@@ -3407,7 +3426,7 @@ static int gfx_v6_0_priv_inst_irq(struct amdgpu_device *adev,
 				  struct amdgpu_iv_entry *entry)
 {
 	DRM_ERROR("Illegal instruction in command stream\n");
-	schedule_work(&adev->reset_work);
+	gfx_v6_0_fault(adev, entry);
 	return 0;
 }
 

drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c

@@ -4959,12 +4959,36 @@ static int gfx_v7_0_eop_irq(struct amdgpu_device *adev,
 	return 0;
 }
 
+static void gfx_v7_0_fault(struct amdgpu_device *adev,
+			   struct amdgpu_iv_entry *entry)
+{
+	struct amdgpu_ring *ring;
+	u8 me_id, pipe_id;
+	int i;
+
+	me_id = (entry->ring_id & 0x0c) >> 2;
+	pipe_id = (entry->ring_id & 0x03) >> 0;
+	switch (me_id) {
+	case 0:
+		drm_sched_fault(&adev->gfx.gfx_ring[0].sched);
+		break;
+	case 1:
+	case 2:
+		for (i = 0; i < adev->gfx.num_compute_rings; i++) {
+			ring = &adev->gfx.compute_ring[i];
+			if ((ring->me == me_id) && (ring->pipe == pipe_id))
+				drm_sched_fault(&ring->sched);
+		}
+		break;
+	}
+}
+
 static int gfx_v7_0_priv_reg_irq(struct amdgpu_device *adev,
 				 struct amdgpu_irq_src *source,
 				 struct amdgpu_iv_entry *entry)
 {
 	DRM_ERROR("Illegal register access in command stream\n");
-	schedule_work(&adev->reset_work);
+	gfx_v7_0_fault(adev, entry);
 	return 0;
 }
 
@@ -4974,7 +4998,7 @@ static int gfx_v7_0_priv_inst_irq(struct amdgpu_device *adev,
 {
 	DRM_ERROR("Illegal instruction in command stream\n");
 	// XXX soft reset the gfx block only
-	schedule_work(&adev->reset_work);
+	gfx_v7_0_fault(adev, entry);
 	return 0;
 }
 

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c

@@ -6738,12 +6738,39 @@ static int gfx_v8_0_eop_irq(struct amdgpu_device *adev,
 	return 0;
 }
 
+static void gfx_v8_0_fault(struct amdgpu_device *adev,
+			   struct amdgpu_iv_entry *entry)
+{
+	u8 me_id, pipe_id, queue_id;
+	struct amdgpu_ring *ring;
+	int i;
+
+	me_id = (entry->ring_id & 0x0c) >> 2;
+	pipe_id = (entry->ring_id & 0x03) >> 0;
+	queue_id = (entry->ring_id & 0x70) >> 4;
+
+	switch (me_id) {
+	case 0:
+		drm_sched_fault(&adev->gfx.gfx_ring[0].sched);
+		break;
+	case 1:
+	case 2:
+		for (i = 0; i < adev->gfx.num_compute_rings; i++) {
+			ring = &adev->gfx.compute_ring[i];
+			if (ring->me == me_id && ring->pipe == pipe_id &&
+			    ring->queue == queue_id)
+				drm_sched_fault(&ring->sched);
+		}
+		break;
+	}
+}
+
 static int gfx_v8_0_priv_reg_irq(struct amdgpu_device *adev,
 				 struct amdgpu_irq_src *source,
 				 struct amdgpu_iv_entry *entry)
 {
 	DRM_ERROR("Illegal register access in command stream\n");
-	schedule_work(&adev->reset_work);
+	gfx_v8_0_fault(adev, entry);
 	return 0;
 }
 
@@ -6752,7 +6779,7 @@ static int gfx_v8_0_priv_inst_irq(struct amdgpu_device *adev,
 				  struct amdgpu_iv_entry *entry)
 {
 	DRM_ERROR("Illegal instruction in command stream\n");
-	schedule_work(&adev->reset_work);
+	gfx_v8_0_fault(adev, entry);
 	return 0;
 }
 

drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c

@@ -4695,12 +4695,39 @@ static int gfx_v9_0_eop_irq(struct amdgpu_device *adev,
 	return 0;
 }
 
+static void gfx_v9_0_fault(struct amdgpu_device *adev,
+			   struct amdgpu_iv_entry *entry)
+{
+	u8 me_id, pipe_id, queue_id;
+	struct amdgpu_ring *ring;
+	int i;
+
+	me_id = (entry->ring_id & 0x0c) >> 2;
+	pipe_id = (entry->ring_id & 0x03) >> 0;
+	queue_id = (entry->ring_id & 0x70) >> 4;
+
+	switch (me_id) {
+	case 0:
+		drm_sched_fault(&adev->gfx.gfx_ring[0].sched);
+		break;
+	case 1:
+	case 2:
+		for (i = 0; i < adev->gfx.num_compute_rings; i++) {
+			ring = &adev->gfx.compute_ring[i];
+			if (ring->me == me_id && ring->pipe == pipe_id &&
+			    ring->queue == queue_id)
+				drm_sched_fault(&ring->sched);
+		}
+		break;
+	}
+}
+
 static int gfx_v9_0_priv_reg_irq(struct amdgpu_device *adev,
 				 struct amdgpu_irq_src *source,
 				 struct amdgpu_iv_entry *entry)
 {
 	DRM_ERROR("Illegal register access in command stream\n");
-	schedule_work(&adev->reset_work);
+	gfx_v9_0_fault(adev, entry);
 	return 0;
 }
 
@@ -4709,7 +4736,7 @@ static int gfx_v9_0_priv_inst_irq(struct amdgpu_device *adev,
 				  struct amdgpu_iv_entry *entry)
 {
 	DRM_ERROR("Illegal instruction in command stream\n");
-	schedule_work(&adev->reset_work);
+	gfx_v9_0_fault(adev, entry);
 	return 0;
 }
 

drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c

@@ -1105,8 +1105,14 @@ static int sdma_v2_4_process_illegal_inst_irq(struct amdgpu_device *adev,
 					      struct amdgpu_irq_src *source,
 					      struct amdgpu_iv_entry *entry)
 {
+	u8 instance_id, queue_id;
+
 	DRM_ERROR("Illegal instruction in SDMA command stream\n");
-	schedule_work(&adev->reset_work);
+	instance_id = (entry->ring_id & 0x3) >> 0;
+	queue_id = (entry->ring_id & 0xc) >> 2;
+
+	if (instance_id <= 1 && queue_id == 0)
+		drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
 	return 0;
 }
 

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

@@ -1440,8 +1440,14 @@ static int sdma_v3_0_process_illegal_inst_irq(struct amdgpu_device *adev,
 					      struct amdgpu_irq_src *source,
 					      struct amdgpu_iv_entry *entry)
 {
+	u8 instance_id, queue_id;
+
 	DRM_ERROR("Illegal instruction in SDMA command stream\n");
-	schedule_work(&adev->reset_work);
+	instance_id = (entry->ring_id & 0x3) >> 0;
+	queue_id = (entry->ring_id & 0xc) >> 2;
+
+	if (instance_id <= 1 && queue_id == 0)
+		drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
 	return 0;
 }
 

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

@@ -1717,12 +1717,29 @@ static int sdma_v4_0_process_illegal_inst_irq(struct amdgpu_device *adev,
 					      struct amdgpu_irq_src *source,
 					      struct amdgpu_iv_entry *entry)
 {
+	int instance;
+
 	DRM_ERROR("Illegal instruction in SDMA command stream\n");
-	schedule_work(&adev->reset_work);
+
+	switch (entry->client_id) {
+	case SOC15_IH_CLIENTID_SDMA0:
+		instance = 0;
+		break;
+	case SOC15_IH_CLIENTID_SDMA1:
+		instance = 1;
+		break;
+	default:
+		return 0;
+	}
+
+	switch (entry->ring_id) {
+	case 0:
+		drm_sched_fault(&adev->sdma.instance[instance].ring.sched);
+		break;
+	}
 	return 0;
 }
 
-
 static void sdma_v4_0_update_medium_grain_clock_gating(
 		struct amdgpu_device *adev,
 		bool enable)