drm/amdkfd: CWSR with software scheduler

This patch adds support to program trap handler settings
when loading driver with software scheduler (sched_policy=2).

Signed-off-by: Mukul Joshi <[e-mail]>
Suggested-by: Jay Cornwall <[e-mail]>
Reviewed-by: Harish Kasiviswanathan <[e-mail]>
Signed-off-by: Alex Deucher <[e-mail]>

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c

···
 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
 		type = RESET_WAVES;
 		break;
+	case KFD_PREEMPT_TYPE_WAVEFRONT_SAVE:
+		type = SAVE_WAVES;
+		break;
 	default:
 		type = DRAIN_PIPE;
 		break;
···
 	adev->gfxhub.funcs->setup_vm_pt_regs(adev, vmid, page_table_base);
 }
 
+static void program_trap_handler_settings(struct kgd_dev *kgd,
+		uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr)
+{
+	struct amdgpu_device *adev = get_amdgpu_device(kgd);
+
+	lock_srbm(kgd, 0, 0, 0, vmid);
+
+	/*
+	 * Program TBA registers
+	 */
+	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TBA_LO),
+			lower_32_bits(tba_addr >> 8));
+	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TBA_HI),
+			upper_32_bits(tba_addr >> 8) |
+			(1 << SQ_SHADER_TBA_HI__TRAP_EN__SHIFT));
+
+	/*
+	 * Program TMA registers
+	 */
+	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TMA_LO),
+			lower_32_bits(tma_addr >> 8));
+	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TMA_HI),
+			upper_32_bits(tma_addr >> 8));
+
+	unlock_srbm(kgd);
+}
+
 const struct kfd2kgd_calls gfx_v10_kfd2kgd = {
 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
···
 	.get_atc_vmid_pasid_mapping_info =
 			get_atc_vmid_pasid_mapping_info,
 	.set_vm_context_page_table_base = set_vm_context_page_table_base,
+	.program_trap_handler_settings = program_trap_handler_settings,
 };

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c

···
 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
 		type = RESET_WAVES;
 		break;
+	case KFD_PREEMPT_TYPE_WAVEFRONT_SAVE:
+		type = SAVE_WAVES;
+		break;
 	default:
 		type = DRAIN_PIPE;
 		break;
···
 
 	/* SDMA is on gfxhub as well for Navi1* series */
 	adev->gfxhub.funcs->setup_vm_pt_regs(adev, vmid, page_table_base);
+}
+
+static void program_trap_handler_settings_v10_3(struct kgd_dev *kgd,
+			uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr)
+{
+	struct amdgpu_device *adev = get_amdgpu_device(kgd);
+
+	lock_srbm(kgd, 0, 0, 0, vmid);
+
+	/*
+	 * Program TBA registers
+	 */
+	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TBA_LO),
+			lower_32_bits(tba_addr >> 8));
+	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TBA_HI),
+			upper_32_bits(tba_addr >> 8) |
+			(1 << SQ_SHADER_TBA_HI__TRAP_EN__SHIFT));
+
+	/*
+	 * Program TMA registers
+	 */
+	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TMA_LO),
+			lower_32_bits(tma_addr >> 8));
+	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TMA_HI),
+			 upper_32_bits(tma_addr >> 8));
+
+	unlock_srbm(kgd);
 }
 
 #if 0
···
 	.address_watch_get_offset = address_watch_get_offset_v10_3,
 	.get_atc_vmid_pasid_mapping_info = NULL,
 	.set_vm_context_page_table_base = set_vm_context_page_table_base_v10_3,
+	.program_trap_handler_settings = program_trap_handler_settings_v10_3,
 #if 0
 	.enable_debug_trap = enable_debug_trap_v10_3,
 	.disable_debug_trap = disable_debug_trap_v10_3,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c

···
 enum hqd_dequeue_request_type {
 	NO_ACTION = 0,
 	DRAIN_PIPE,
-	RESET_WAVES
+	RESET_WAVES,
+	SAVE_WAVES
 };
 
 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
···
 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
 		type = RESET_WAVES;
 		break;
+	case KFD_PREEMPT_TYPE_WAVEFRONT_SAVE:
+		type = SAVE_WAVES;
+		break;
 	default:
 		type = DRAIN_PIPE;
 		break;
···
 				adev->gfx.cu_info.max_waves_per_simd;
 }
 
+static void kgd_gfx_v9_program_trap_handler_settings(struct kgd_dev *kgd,
+                        uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr)
+{
+	struct amdgpu_device *adev = get_amdgpu_device(kgd);
+
+	lock_srbm(kgd, 0, 0, 0, vmid);
+
+	/*
+	 * Program TBA registers
+	 */
+	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TBA_LO),
+                        lower_32_bits(tba_addr >> 8));
+	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TBA_HI),
+                        upper_32_bits(tba_addr >> 8));
+
+	/*
+	 * Program TMA registers
+	 */
+	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TMA_LO),
+			lower_32_bits(tma_addr >> 8));
+	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TMA_HI),
+			upper_32_bits(tma_addr >> 8));
+
+	unlock_srbm(kgd);
+}
+
 const struct kfd2kgd_calls gfx_v9_kfd2kgd = {
 	.program_sh_mem_settings = kgd_gfx_v9_program_sh_mem_settings,
 	.set_pasid_vmid_mapping = kgd_gfx_v9_set_pasid_vmid_mapping,
···
 			kgd_gfx_v9_get_atc_vmid_pasid_mapping_info,
 	.set_vm_context_page_table_base = kgd_gfx_v9_set_vm_context_page_table_base,
 	.get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy,
+	.program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings,
 };

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c

···
 	WARN_ON(!old);
 }
 
+static void program_trap_handler_settings(struct device_queue_manager *dqm,
+				struct qcm_process_device *qpd)
+{
+	if (dqm->dev->kfd2kgd->program_trap_handler_settings)
+		dqm->dev->kfd2kgd->program_trap_handler_settings(
+						dqm->dev->kgd, qpd->vmid,
+						qpd->tba_addr, qpd->tma_addr);
+}
+
 static int allocate_vmid(struct device_queue_manager *dqm,
 			struct qcm_process_device *qpd,
 			struct queue *q)
···
 	q->properties.vmid = allocated_vmid;
 
 	program_sh_mem_settings(dqm, qpd);
+
+	if (dqm->dev->device_info->asic_family >= CHIP_VEGA10 &&
+	    dqm->dev->cwsr_enabled)
+		program_trap_handler_settings(dqm, qpd);
 
 	/* qpd->page_table_base is set earlier when register_process()
 	 * is called, i.e. when the first queue is created.
···
 		}
 
 		retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
-				KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN,
+				(dqm->dev->cwsr_enabled?
+				 KFD_PREEMPT_TYPE_WAVEFRONT_SAVE:
+				KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN),
 				KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
 		if (retval) {
 			pr_err("destroy mqd failed\n");
···
 			continue;
 
 		retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
-				KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN,
+				(dqm->dev->cwsr_enabled?
+				 KFD_PREEMPT_TYPE_WAVEFRONT_SAVE:
+				KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN),
 				KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
 		if (retval && !ret)
 			/* Return the first error, but keep going to

drivers/gpu/drm/amd/include/kgd_kfd_interface.h

···
 enum kfd_preempt_type {
 	KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN = 0,
 	KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
+	KFD_PREEMPT_TYPE_WAVEFRONT_SAVE
 };
 
 struct kfd_vm_fault_info {
···
 
 	void (*get_cu_occupancy)(struct kgd_dev *kgd, int pasid, int *wave_cnt,
 			int *max_waves_per_cu);
+	void (*program_trap_handler_settings)(struct kgd_dev *kgd,
+			uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr);
 };
 
 #endif	/* KGD_KFD_INTERFACE_H_INCLUDED */