drm/amdkfd: Move TLB flushing logic into amdgpu

This will make it possible for amdgpu GEM ioctls to flush TLBs on compute
VMs.

This removes VMID-based TLB flushing and always uses PASID-based
flushing. This still works because it scans the VMID-PASID mapping
registers to find the right VMID. It's only slightly less efficient. This
is not a production use case.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

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

@@ -710,35 +710,6 @@ bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid)
 	return false;
 }
 
-int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct amdgpu_device *adev,
-				     uint16_t vmid)
-{
-	if (adev->family == AMDGPU_FAMILY_AI) {
-		int i;
-
-		for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS)
-			amdgpu_gmc_flush_gpu_tlb(adev, vmid, i, 0);
-	} else {
-		amdgpu_gmc_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB(0), 0);
-	}
-
-	return 0;
-}
-
-int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
-				      uint16_t pasid,
-				      enum TLB_FLUSH_TYPE flush_type,
-				      uint32_t inst)
-{
-	bool all_hub = false;
-
-	if (adev->family == AMDGPU_FAMILY_AI ||
-	    adev->family == AMDGPU_FAMILY_RV)
-		all_hub = true;
-
-	return amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, flush_type, all_hub, inst);
-}
-
 bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev)
 {
 	return adev->have_atomics_support;

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

@@ -162,11 +162,6 @@ int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev,
 				uint32_t *ib_cmd, uint32_t ib_len);
 void amdgpu_amdkfd_set_compute_idle(struct amdgpu_device *adev, bool idle);
 bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev);
-int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct amdgpu_device *adev,
-				uint16_t vmid);
-int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
-				uint16_t pasid, enum TLB_FLUSH_TYPE flush_type,
-				uint32_t inst);
 
 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid);
 

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

@@ -1437,6 +1437,50 @@ int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
 	return 0;
 }
 
+/**
+ * amdgpu_vm_flush_compute_tlb - Flush TLB on compute VM
+ *
+ * @adev: amdgpu_device pointer
+ * @vm: requested vm
+ * @flush_type: flush type
+ *
+ * Flush TLB if needed for a compute VM.
+ *
+ * Returns:
+ * 0 for success.
+ */
+int amdgpu_vm_flush_compute_tlb(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm,
+				uint32_t flush_type,
+				uint32_t xcc_mask)
+{
+	uint64_t tlb_seq = amdgpu_vm_tlb_seq(vm);
+	bool all_hub = false;
+	int xcc = 0, r = 0;
+
+	WARN_ON_ONCE(!vm->is_compute_context);
+
+	/*
+	 * It can be that we race and lose here, but that is extremely unlikely
+	 * and the worst thing which could happen is that we flush the changes
+	 * into the TLB once more which is harmless.
+	 */
+	if (atomic64_xchg(&vm->kfd_last_flushed_seq, tlb_seq) == tlb_seq)
+		return 0;
+
+	if (adev->family == AMDGPU_FAMILY_AI ||
+	    adev->family == AMDGPU_FAMILY_RV)
+		all_hub = true;
+
+	for_each_inst(xcc, xcc_mask) {
+		r = amdgpu_gmc_flush_gpu_tlb_pasid(adev, vm->pasid, flush_type,
+						   all_hub, xcc);
+		if (r)
+			break;
+	}
+	return r;
+}
+
 /**
  * amdgpu_vm_bo_add - add a bo to a specific vm
  *

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

@@ -324,6 +324,7 @@ struct amdgpu_vm {
 	/* Last finished delayed update */
 	atomic64_t		tlb_seq;
 	struct dma_fence	*last_tlb_flush;
+	atomic64_t		kfd_last_flushed_seq;
 
 	/* How many times we had to re-generate the page tables */
 	uint64_t		generation;
@@ -445,6 +446,10 @@ int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
 			   struct amdgpu_vm *vm,
 			   struct ww_acquire_ctx *ticket);
+int amdgpu_vm_flush_compute_tlb(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm,
+				uint32_t flush_type,
+				uint32_t xcc_mask);
 void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
 			    struct amdgpu_vm *vm, struct amdgpu_bo *bo);
 int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,

drivers/gpu/drm/amd/amdkfd/kfd_priv.h

@@ -748,7 +748,6 @@ struct kfd_process_device {
 	/* VM context for GPUVM allocations */
 	struct file *drm_file;
 	void *drm_priv;
-	atomic64_t tlb_seq;
 
 	/* GPUVM allocations storage */
 	struct idr alloc_idr;
@@ -1462,7 +1461,14 @@ void kfd_signal_reset_event(struct kfd_node *dev);
 
 void kfd_signal_poison_consumed_event(struct kfd_node *dev, u32 pasid);
 
-void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type);
+static inline void kfd_flush_tlb(struct kfd_process_device *pdd,
+				 enum TLB_FLUSH_TYPE type)
+{
+	struct amdgpu_device *adev = pdd->dev->adev;
+	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
+
+	amdgpu_vm_flush_compute_tlb(adev, vm, type, pdd->dev->xcc_mask);
+}
 
 static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev)
 {

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

@@ -1667,7 +1667,6 @@ int kfd_process_device_init_vm(struct kfd_process_device *pdd,
 		return ret;
 	}
 	pdd->drm_priv = drm_file->private_data;
-	atomic64_set(&pdd->tlb_seq, 0);
 
 	ret = kfd_process_device_reserve_ib_mem(pdd);
 	if (ret)
@@ -2059,36 +2058,6 @@ int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process,
 			       KFD_CWSR_TBA_TMA_SIZE, vma->vm_page_prot);
 }
 
-void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type)
-{
-	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
-	uint64_t tlb_seq = amdgpu_vm_tlb_seq(vm);
-	struct kfd_node *dev = pdd->dev;
-	uint32_t xcc_mask = dev->xcc_mask;
-	int xcc = 0;
-
-	/*
-	 * It can be that we race and lose here, but that is extremely unlikely
-	 * and the worst thing which could happen is that we flush the changes
-	 * into the TLB once more which is harmless.
-	 */
-	if (atomic64_xchg(&pdd->tlb_seq, tlb_seq) == tlb_seq)
-		return;
-
-	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
-		/* Nothing to flush until a VMID is assigned, which
-		 * only happens when the first queue is created.
-		 */
-		if (pdd->qpd.vmid)
-			amdgpu_amdkfd_flush_gpu_tlb_vmid(dev->adev,
-							pdd->qpd.vmid);
-	} else {
-		for_each_inst(xcc, xcc_mask)
-			amdgpu_amdkfd_flush_gpu_tlb_pasid(
-				dev->adev, pdd->process->pasid, type, xcc);
-	}
-}
-
 /* assumes caller holds process lock. */
 int kfd_process_drain_interrupts(struct kfd_process_device *pdd)
 {