drm/amdkfd: pass kfd_node ref to svm migration api

This work is required for GC 9.4.3, previous to support memory
partitions per node at SVM. When multiple partition is configured,
every BO should be allocated inside one specific partition which
corresponds to the current amdgpu_device and kfd_node.

v2: squash in compilation fix (Alex)
v3: squash in fix for pre-gfx 9.4.3 (Alex)
v4: squash in best_loc fix (Alex)
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

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

@@ -2441,7 +2441,8 @@ void amdgpu_vm_set_task_info(struct amdgpu_vm *vm)
  * shouldn't be reported any more.
  */
 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
-			    uint64_t addr, bool write_fault)
+			    u32 client_id, u32 node_id, uint64_t addr,
+			    bool write_fault)
 {
 	bool is_compute_context = false;
 	struct amdgpu_bo *root;
@@ -2465,8 +2466,8 @@ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
 
 	addr /= AMDGPU_GPU_PAGE_SIZE;
 
-	if (is_compute_context &&
-	    !svm_range_restore_pages(adev, pasid, addr, write_fault)) {
+	if (is_compute_context && !svm_range_restore_pages(adev, pasid, client_id,
+	    node_id, addr, write_fault)) {
 		amdgpu_bo_unref(&root);
 		return true;
 	}

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

@@ -455,7 +455,8 @@ void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
 void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid,
 			     struct amdgpu_task_info *task_info);
 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
-			    uint64_t addr, bool write_fault);
+			    u32 client_id, u32 node_id, uint64_t addr,
+			    bool write_fault);
 
 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
 

drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c

@@ -139,7 +139,7 @@ static int gmc_v10_0_process_interrupt(struct amdgpu_device *adev,
 		/* Try to handle the recoverable page faults by filling page
 		 * tables
 		 */
-		if (amdgpu_vm_handle_fault(adev, entry->pasid, addr, write_fault))
+		if (amdgpu_vm_handle_fault(adev, entry->pasid, 0, 0, addr, write_fault))
 			return 1;
 	}
 

drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c

@@ -557,11 +557,24 @@ static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
 	u64 addr;
 	uint32_t cam_index = 0;
 	int ret;
-	uint32_t node_id;
+	uint32_t node_id = 0;
 
 	addr = (u64)entry->src_data[0] << 12;
 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
 
+	if (entry->client_id == SOC15_IH_CLIENTID_VMC) {
+		hub_name = "mmhub0";
+		hub = &adev->vmhub[AMDGPU_MMHUB0(0)];
+	} else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) {
+		hub_name = "mmhub1";
+		hub = &adev->vmhub[AMDGPU_MMHUB1(0)];
+	} else {
+		hub_name = "gfxhub0";
+		node_id = (adev->ip_versions[GC_HWIP][0] ==
+			   IP_VERSION(9, 4, 3)) ? entry->node_id : 0;
+		hub = &adev->vmhub[node_id/2];
+	}
+
 	if (retry_fault) {
 		if (adev->irq.retry_cam_enabled) {
 			/* Delegate it to a different ring if the hardware hasn't
@@ -574,7 +587,8 @@ static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
 
 			cam_index = entry->src_data[2] & 0x3ff;
 
-			ret = amdgpu_vm_handle_fault(adev, entry->pasid, addr, write_fault);
+			ret = amdgpu_vm_handle_fault(adev, entry->pasid, entry->client_id, node_id,
+						     addr, write_fault);
 			WDOORBELL32(adev->irq.retry_cam_doorbell_index, cam_index);
 			if (ret)
 				return 1;
@@ -596,7 +610,8 @@ static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
 			/* Try to handle the recoverable page faults by filling page
 			 * tables
 			 */
-			if (amdgpu_vm_handle_fault(adev, entry->pasid, addr, write_fault))
+			if (amdgpu_vm_handle_fault(adev, entry->pasid, entry->client_id, node_id,
+						   addr, write_fault))
 				return 1;
 		}
 	}
@@ -604,18 +619,6 @@ static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
 	if (!printk_ratelimit())
 		return 0;
 
-	if (entry->client_id == SOC15_IH_CLIENTID_VMC) {
-		hub_name = "mmhub0";
-		hub = &adev->vmhub[AMDGPU_MMHUB0(0)];
-	} else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) {
-		hub_name = "mmhub1";
-		hub = &adev->vmhub[AMDGPU_MMHUB1(0)];
-	} else {
-		hub_name = "gfxhub0";
-		node_id = (adev->ip_versions[GC_HWIP][0] ==
-			   IP_VERSION(9, 4, 3)) ? entry->node_id : 0;
-		hub = &adev->vmhub[node_id/2];
-	}
 
 	memset(&task_info, 0, sizeof(struct amdgpu_task_info));
 	amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);

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

@@ -287,11 +287,12 @@ static unsigned long svm_migrate_unsuccessful_pages(struct migrate_vma *migrate)
 }
 
 static int
-svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
+svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
 			 struct migrate_vma *migrate, struct dma_fence **mfence,
 			 dma_addr_t *scratch, uint64_t ttm_res_offset)
 {
-	uint64_t npages = migrate->npages;
+	uint64_t npages = migrate->cpages;
+	struct amdgpu_device *adev = node->adev;
 	struct device *dev = adev->dev;
 	struct amdgpu_res_cursor cursor;
 	dma_addr_t *src;
@@ -321,7 +322,7 @@ svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 					      DMA_TO_DEVICE);
 			r = dma_mapping_error(dev, src[i]);
 			if (r) {
-				dev_err(adev->dev, "%s: fail %d dma_map_page\n",
+				dev_err(dev, "%s: fail %d dma_map_page\n",
 					__func__, r);
 				goto out_free_vram_pages;
 			}
@@ -390,12 +391,13 @@ svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 }
 
 static long
-svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
+svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
 			struct vm_area_struct *vma, uint64_t start,
 			uint64_t end, uint32_t trigger, uint64_t ttm_res_offset)
 {
 	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
 	uint64_t npages = (end - start) >> PAGE_SHIFT;
+	struct amdgpu_device *adev = node->adev;
 	struct kfd_process_device *pdd;
 	struct dma_fence *mfence = NULL;
 	struct migrate_vma migrate = { 0 };
@@ -445,7 +447,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 	else
 		pr_debug("0x%lx pages migrated\n", cpages);
 
-	r = svm_migrate_copy_to_vram(adev, prange, &migrate, &mfence, scratch, ttm_res_offset);
+	r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset);
 	migrate_vma_pages(&migrate);
 
 	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
@@ -465,7 +467,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 	kvfree(buf);
 out:
 	if (!r && cpages) {
-		pdd = svm_range_get_pdd_by_adev(prange, adev);
+		pdd = svm_range_get_pdd_by_node(prange, node);
 		if (pdd)
 			WRITE_ONCE(pdd->page_in, pdd->page_in + cpages);
 
@@ -492,8 +494,8 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc,
 {
 	unsigned long addr, start, end;
 	struct vm_area_struct *vma;
-	struct amdgpu_device *adev;
 	uint64_t ttm_res_offset;
+	struct kfd_node *node;
 	unsigned long cpages = 0;
 	long r = 0;
 
@@ -503,9 +505,9 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc,
 		return 0;
 	}
 
-	adev = svm_range_get_adev_by_id(prange, best_loc);
-	if (!adev) {
-		pr_debug("failed to get device by id 0x%x\n", best_loc);
+	node = svm_range_get_node_by_id(prange, best_loc);
+	if (!node) {
+		pr_debug("failed to get kfd node by id 0x%x\n", best_loc);
 		return -ENODEV;
 	}
 
@@ -515,9 +517,9 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc,
 	start = prange->start << PAGE_SHIFT;
 	end = (prange->last + 1) << PAGE_SHIFT;
 
-	r = svm_range_vram_node_new(adev, prange, true);
+	r = svm_range_vram_node_new(node, prange, true);
 	if (r) {
-		dev_dbg(adev->dev, "fail %ld to alloc vram\n", r);
+		dev_dbg(node->adev->dev, "fail %ld to alloc vram\n", r);
 		return r;
 	}
 	ttm_res_offset = prange->offset << PAGE_SHIFT;
@@ -530,7 +532,7 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc,
 			break;
 
 		next = min(vma->vm_end, end);
-		r = svm_migrate_vma_to_vram(adev, prange, vma, addr, next, trigger, ttm_res_offset);
+		r = svm_migrate_vma_to_vram(node, prange, vma, addr, next, trigger, ttm_res_offset);
 		if (r < 0) {
 			pr_debug("failed %ld to migrate\n", r);
 			break;
@@ -663,7 +665,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
  *   positive values - partial migration, number of pages not migrated
  */
 static long
-svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
+svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 		       struct vm_area_struct *vma, uint64_t start, uint64_t end,
 		       uint32_t trigger, struct page *fault_page)
 {
@@ -671,6 +673,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 	uint64_t npages = (end - start) >> PAGE_SHIFT;
 	unsigned long upages = npages;
 	unsigned long cpages = 0;
+	struct amdgpu_device *adev = node->adev;
 	struct kfd_process_device *pdd;
 	struct dma_fence *mfence = NULL;
 	struct migrate_vma migrate = { 0 };
@@ -745,7 +748,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 	kvfree(buf);
 out:
 	if (!r && cpages) {
-		pdd = svm_range_get_pdd_by_adev(prange, adev);
+		pdd = svm_range_get_pdd_by_node(prange, node);
 		if (pdd)
 			WRITE_ONCE(pdd->page_out, pdd->page_out + cpages);
 	}
@@ -766,7 +769,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm,
 			    uint32_t trigger, struct page *fault_page)
 {
-	struct amdgpu_device *adev;
+	struct kfd_node *node;
 	struct vm_area_struct *vma;
 	unsigned long addr;
 	unsigned long start;
@@ -780,13 +783,11 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm,
 		return 0;
 	}
 
-	adev = svm_range_get_adev_by_id(prange, prange->actual_loc);
-	if (!adev) {
-		pr_debug("failed to get device by id 0x%x\n",
-			 prange->actual_loc);
+	node = svm_range_get_node_by_id(prange, prange->actual_loc);
+	if (!node) {
+		pr_debug("failed to get kfd node by id 0x%x\n", prange->actual_loc);
 		return -ENODEV;
 	}
-
 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] from gpu 0x%x to ram\n",
 		 prange->svms, prange, prange->start, prange->last,
 		 prange->actual_loc);
@@ -805,7 +806,7 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm,
 		}
 
 		next = min(vma->vm_end, end);
-		r = svm_migrate_vma_to_ram(adev, prange, vma, addr, next, trigger,
+		r = svm_migrate_vma_to_ram(node, prange, vma, addr, next, trigger,
 			fault_page);
 		if (r < 0) {
 			pr_debug("failed %ld to migrate prange %p\n", r, prange);

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

@@ -289,6 +289,7 @@ struct kfd_node {
 	 * from the HW ring into a SW ring.
 	 */
 	bool interrupts_active;
+	uint32_t interrupt_bitmap; /* Only used for GFX 9.4.3 */
 
 	/* QCM Device instance */
 	struct device_queue_manager *dqm;
@@ -971,9 +972,8 @@ struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid);
 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
 
 int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id);
-int kfd_process_gpuid_from_adev(struct kfd_process *p,
-			       struct amdgpu_device *adev, uint32_t *gpuid,
-			       uint32_t *gpuidx);
+int kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,
+				uint32_t *gpuid, uint32_t *gpuidx);
 static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p,
 				uint32_t gpuidx, uint32_t *gpuid) {
 	return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL;
@@ -1073,6 +1073,30 @@ struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
 struct kfd_node *kfd_device_by_id(uint32_t gpu_id);
 struct kfd_node *kfd_device_by_pci_dev(const struct pci_dev *pdev);
 struct kfd_node *kfd_device_by_adev(const struct amdgpu_device *adev);
+static inline bool kfd_irq_is_from_node(struct kfd_node *node, uint32_t client_id,
+				     uint32_t node_id)
+{
+	if ((node->interrupt_bitmap & (0x1U << node_id)) ||
+	    ((node_id % 4) == 0 &&
+	    (node->interrupt_bitmap >> 16) & (0x1U << client_id)))
+		return true;
+
+	return false;
+}
+static inline struct kfd_node *kfd_node_by_irq_ids(struct amdgpu_device *adev,
+					uint32_t client_id, uint32_t node_id) {
+	struct kfd_dev *dev = adev->kfd.dev;
+	uint32_t i;
+
+	if (adev->ip_versions[GC_HWIP][0] != IP_VERSION(9, 4, 3))
+		return dev->nodes[0];
+
+	for (i = 0; i < dev->num_nodes; i++)
+		if (kfd_irq_is_from_node(dev->nodes[i], client_id, node_id))
+			return dev->nodes[i];
+
+	return NULL;
+}
 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev);
 int kfd_numa_node_to_apic_id(int numa_node_id);
 void kfd_double_confirm_iommu_support(struct kfd_dev *gpu);

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

@@ -1891,13 +1891,13 @@ int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id)
 }
 
 int
-kfd_process_gpuid_from_adev(struct kfd_process *p, struct amdgpu_device *adev,
-			   uint32_t *gpuid, uint32_t *gpuidx)
+kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,
+			    uint32_t *gpuid, uint32_t *gpuidx)
 {
 	int i;
 
 	for (i = 0; i < p->n_pdds; i++)
-		if (p->pdds[i] && p->pdds[i]->dev->adev == adev) {
+		if (p->pdds[i] && p->pdds[i]->dev == node) {
 			*gpuid = p->pdds[i]->user_gpu_id;
 			*gpuidx = i;
 			return 0;

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

@@ -48,6 +48,7 @@ struct svm_range_bo {
 	struct work_struct		eviction_work;
 	uint32_t			evicting;
 	struct work_struct		release_work;
+	struct kfd_node			*node;
 };
 
 enum svm_work_list_ops {
@@ -163,16 +164,17 @@ int svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
 struct svm_range *svm_range_from_addr(struct svm_range_list *svms,
 				      unsigned long addr,
 				      struct svm_range **parent);
-struct amdgpu_device *svm_range_get_adev_by_id(struct svm_range *prange,
-					       uint32_t id);
-int svm_range_vram_node_new(struct amdgpu_device *adev,
-			    struct svm_range *prange, bool clear);
+struct kfd_node *svm_range_get_node_by_id(struct svm_range *prange,
+					  uint32_t gpu_id);
+int svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange,
+			    bool clear);
 void svm_range_vram_node_free(struct svm_range *prange);
 int svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
 			       unsigned long addr, struct svm_range *parent,
 			       struct svm_range *prange);
-int svm_range_restore_pages(struct amdgpu_device *adev,
-			    unsigned int pasid, uint64_t addr, bool write_fault);
+int svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
+			    uint32_t client_id, uint32_t node_id, uint64_t addr,
+			    bool write_fault);
 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence);
 void svm_range_add_list_work(struct svm_range_list *svms,
 			     struct svm_range *prange, struct mm_struct *mm,
@@ -192,7 +194,7 @@ int kfd_criu_restore_svm(struct kfd_process *p,
 			 uint64_t max_priv_data_size);
 int kfd_criu_resume_svm(struct kfd_process *p);
 struct kfd_process_device *
-svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev);
+svm_range_get_pdd_by_node(struct svm_range *prange, struct kfd_node *node);
 void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_struct *mm);
 
 /* SVM API and HMM page migration work together, device memory type
@@ -219,8 +221,9 @@ static inline void svm_range_list_fini(struct kfd_process *p)
 }
 
 static inline int svm_range_restore_pages(struct amdgpu_device *adev,
-					  unsigned int pasid, uint64_t addr,
-					  bool write_fault)
+					  unsigned int pasid,
+					  uint32_t client_id, uint32_t node_id,
+					  uint64_t addr, bool write_fault)
 {
 	return -EFAULT;
 }