drm/amdkfd: track unified memory reservation with xnack off

[WHY]
Unified memory with xnack off should be tracked, as userptr mappings
and legacy allocations do. To avoid oversuscribe system memory when
xnack off.
[How]
Exposing functions reserve_mem_limit and unreserve_mem_limit to SVM
API and call them on every prange creation and free.
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_amdkfd.h

@@ -305,6 +305,10 @@ bool amdgpu_amdkfd_bo_mapped_to_dev(struct amdgpu_device *adev, struct kgd_mem *
 void amdgpu_amdkfd_block_mmu_notifications(void *p);
 int amdgpu_amdkfd_criu_resume(void *p);
 bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev);
+int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
+		uint64_t size, u32 alloc_flag);
+void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
+		uint64_t size, u32 alloc_flag);
 
 #if IS_ENABLED(CONFIG_HSA_AMD)
 void amdgpu_amdkfd_gpuvm_init_mem_limits(void);

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

@@ -129,7 +129,7 @@ void amdgpu_amdkfd_reserve_system_mem(uint64_t size)
  *
  * Return: returns -ENOMEM in case of error, ZERO otherwise
  */
-static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
+int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 		uint64_t size, u32 alloc_flag)
 {
 	uint64_t reserved_for_pt =
@@ -169,7 +169,7 @@ static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 	     kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) ||
 	    (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
 	     kfd_mem_limit.max_ttm_mem_limit) ||
-	    (adev->kfd.vram_used + vram_needed >
+	    (adev && adev->kfd.vram_used + vram_needed >
 	     adev->gmc.real_vram_size -
 	     atomic64_read(&adev->vram_pin_size) -
 	     reserved_for_pt)) {
@@ -180,7 +180,10 @@ static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 	/* Update memory accounting by decreasing available system
 	 * memory, TTM memory and GPU memory as computed above
 	 */
-	adev->kfd.vram_used += vram_needed;
+	WARN_ONCE(vram_needed && !adev,
+		  "adev reference can't be null when vram is used");
+	if (adev)
+		adev->kfd.vram_used += vram_needed;
 	kfd_mem_limit.system_mem_used += system_mem_needed;
 	kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
 
@@ -189,7 +192,7 @@ static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 	return ret;
 }
 
-static void unreserve_mem_limit(struct amdgpu_device *adev,
+void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
 		uint64_t size, u32 alloc_flag)
 {
 	spin_lock(&kfd_mem_limit.mem_limit_lock);
@@ -198,7 +201,10 @@ static void unreserve_mem_limit(struct amdgpu_device *adev,
 		kfd_mem_limit.system_mem_used -= size;
 		kfd_mem_limit.ttm_mem_used -= size;
 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
-		adev->kfd.vram_used -= ALIGN(size, VRAM_ALLOCATION_ALIGN);
+		WARN_ONCE(!adev,
+			  "adev reference can't be null when alloc mem flags vram is set");
+		if (adev)
+			adev->kfd.vram_used -= ALIGN(size, VRAM_ALLOCATION_ALIGN);
 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
 		kfd_mem_limit.system_mem_used -= size;
 	} else if (!(alloc_flag &
@@ -207,8 +213,7 @@ static void unreserve_mem_limit(struct amdgpu_device *adev,
 		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
 		goto release;
 	}
-
-	WARN_ONCE(adev->kfd.vram_used < 0,
+	WARN_ONCE(adev && adev->kfd.vram_used < 0,
 		  "KFD VRAM memory accounting unbalanced");
 	WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
 		  "KFD TTM memory accounting unbalanced");
@@ -225,7 +230,7 @@ void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo)
 	u32 alloc_flags = bo->kfd_bo->alloc_flags;
 	u64 size = amdgpu_bo_size(bo);
 
-	unreserve_mem_limit(adev, size, alloc_flags);
+	amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags);
 
 	kfree(bo->kfd_bo);
 }
@@ -1749,7 +1754,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	/* Don't unreserve system mem limit twice */
 	goto err_reserve_limit;
 err_bo_create:
-	unreserve_mem_limit(adev, size, flags);
+	amdgpu_amdkfd_unreserve_mem_limit(adev, size, flags);
 err_reserve_limit:
 	mutex_destroy(&(*mem)->lock);
 	if (gobj)

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

@@ -266,13 +266,22 @@ void svm_range_free_dma_mappings(struct svm_range *prange)
 	}
 }
 
-static void svm_range_free(struct svm_range *prange)
+static void svm_range_free(struct svm_range *prange, bool update_mem_usage)
 {
+	uint64_t size = (prange->last - prange->start + 1) << PAGE_SHIFT;
+	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
+
 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange,
 		 prange->start, prange->last);
 
 	svm_range_vram_node_free(prange);
 	svm_range_free_dma_mappings(prange);
+
+	if (update_mem_usage && !p->xnack_enabled) {
+		pr_debug("unreserve mem limit: %lld\n", size);
+		amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
+					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR);
+	}
 	mutex_destroy(&prange->lock);
 	mutex_destroy(&prange->migrate_mutex);
 	kfree(prange);
@@ -291,7 +300,7 @@ svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc,
 
 static struct
 svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
-			 uint64_t last)
+			 uint64_t last, bool update_mem_usage)
 {
 	uint64_t size = last - start + 1;
 	struct svm_range *prange;
@@ -300,6 +309,15 @@ svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
 	prange = kzalloc(sizeof(*prange), GFP_KERNEL);
 	if (!prange)
 		return NULL;
+
+	p = container_of(svms, struct kfd_process, svms);
+	if (!p->xnack_enabled && update_mem_usage &&
+	    amdgpu_amdkfd_reserve_mem_limit(NULL, size << PAGE_SHIFT,
+					    KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)) {
+		pr_info("SVM mapping failed, exceeds resident system memory limit\n");
+		kfree(prange);
+		return NULL;
+	}
 	prange->npages = size;
 	prange->svms = svms;
 	prange->start = start;
@@ -314,7 +332,6 @@ svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
 	mutex_init(&prange->migrate_mutex);
 	mutex_init(&prange->lock);
 
-	p = container_of(svms, struct kfd_process, svms);
 	if (p->xnack_enabled)
 		bitmap_copy(prange->bitmap_access, svms->bitmap_supported,
 			    MAX_GPU_INSTANCE);
@@ -1007,9 +1024,9 @@ svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
 
 	svms = prange->svms;
 	if (old_start == start)
-		*new = svm_range_new(svms, last + 1, old_last);
+		*new = svm_range_new(svms, last + 1, old_last, false);
 	else
-		*new = svm_range_new(svms, old_start, start - 1);
+		*new = svm_range_new(svms, old_start, start - 1, false);
 	if (!*new)
 		return -ENOMEM;
 
@@ -1017,7 +1034,7 @@ svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
 	if (r) {
 		pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n",
 			 r, old_start, old_last, start, last);
-		svm_range_free(*new);
+		svm_range_free(*new, false);
 		*new = NULL;
 	}
 
@@ -1852,7 +1869,7 @@ static struct svm_range *svm_range_clone(struct svm_range *old)
 {
 	struct svm_range *new;
 
-	new = svm_range_new(old->svms, old->start, old->last);
+	new = svm_range_new(old->svms, old->start, old->last, false);
 	if (!new)
 		return NULL;
 
@@ -1905,7 +1922,7 @@ svm_range_split_new(struct svm_range_list *svms, uint64_t start, uint64_t last,
 	while (last >= start) {
 		l = min(last, ALIGN_DOWN(start + max_pages, max_pages) - 1);
 
-		prange = svm_range_new(svms, start, l);
+		prange = svm_range_new(svms, start, l, true);
 		if (!prange)
 			return -ENOMEM;
 		list_add(&prange->list, insert_list);
@@ -1956,6 +1973,7 @@ svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
 	struct interval_tree_node *node;
 	struct svm_range *prange;
 	struct svm_range *tmp;
+	struct list_head new_list;
 	int r = 0;
 
 	pr_debug("svms 0x%p [0x%llx 0x%lx]\n", &p->svms, start, last);
@@ -1963,6 +1981,7 @@ svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
 	INIT_LIST_HEAD(update_list);
 	INIT_LIST_HEAD(insert_list);
 	INIT_LIST_HEAD(remove_list);
+	INIT_LIST_HEAD(&new_list);
 
 	node = interval_tree_iter_first(&svms->objects, start, last);
 	while (node) {
@@ -2020,7 +2039,7 @@ svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
 		if (node->start > start) {
 			r = svm_range_split_new(svms, start, node->start - 1,
 						READ_ONCE(max_svm_range_pages),
-						insert_list, update_list);
+						&new_list, update_list);
 			if (r)
 				goto out;
 		}
@@ -2033,12 +2052,17 @@ svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
 	if (start <= last)
 		r = svm_range_split_new(svms, start, last,
 					READ_ONCE(max_svm_range_pages),
-					insert_list, update_list);
+					&new_list, update_list);
 
 out:
-	if (r)
+	if (r) {
 		list_for_each_entry_safe(prange, tmp, insert_list, list)
-			svm_range_free(prange);
+			svm_range_free(prange, false);
+		list_for_each_entry_safe(prange, tmp, &new_list, list)
+			svm_range_free(prange, true);
+	} else {
+		list_splice(&new_list, insert_list);
+	}
 
 	return r;
 }
@@ -2085,7 +2109,7 @@ svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange,
 			 svms, prange, prange->start, prange->last);
 		svm_range_unlink(prange);
 		svm_range_remove_notifier(prange);
-		svm_range_free(prange);
+		svm_range_free(prange, true);
 		break;
 	case SVM_OP_UPDATE_RANGE_NOTIFIER:
 		pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n",
@@ -2648,14 +2672,14 @@ svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev,
 		last = addr;
 	}
 
-	prange = svm_range_new(&p->svms, start, last);
+	prange = svm_range_new(&p->svms, start, last, true);
 	if (!prange) {
 		pr_debug("Failed to create prange in address [0x%llx]\n", addr);
 		return NULL;
 	}
 	if (kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpuidx)) {
 		pr_debug("failed to get gpuid from kgd\n");
-		svm_range_free(prange);
+		svm_range_free(prange, true);
 		return NULL;
 	}
 
@@ -2955,7 +2979,7 @@ void svm_range_list_fini(struct kfd_process *p)
 	list_for_each_entry_safe(prange, next, &p->svms.list, list) {
 		svm_range_unlink(prange);
 		svm_range_remove_notifier(prange);
-		svm_range_free(prange);
+		svm_range_free(prange, true);
 	}
 
 	mutex_destroy(&p->svms.lock);
@@ -3371,7 +3395,7 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
 			 prange->last);
 		svm_range_unlink(prange);
 		svm_range_remove_notifier(prange);
-		svm_range_free(prange);
+		svm_range_free(prange, false);
 	}
 
 	mmap_write_downgrade(mm);