drm/amdkfd: Track GPU memory utilization per process

Track GPU VRAM usage on a per process basis and report it through
sysfs.

Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

5 files changed, 84 insertions, 12 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index d065c50582eb..a501026e829c 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -65,6 +65,7 @@ struct kgd_mem {
 	struct amdgpu_sync sync;
 
 	bool aql_queue;
+	bool is_imported;
 };
 
 /* KFD Memory Eviction */
@@ -219,7 +220,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 		void *vm, struct kgd_mem **mem,
 		uint64_t *offset, uint32_t flags);
 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
-		struct kgd_dev *kgd, struct kgd_mem *mem);
+		struct kgd_dev *kgd, struct kgd_mem *mem, uint64_t *size);
 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
 		struct kgd_dev *kgd, struct kgd_mem *mem, void *vm);
 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index 0768b7eb7683..1247938b1ec1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -1277,7 +1277,7 @@ err:
 }
 
 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
-		struct kgd_dev *kgd, struct kgd_mem *mem)
+		struct kgd_dev *kgd, struct kgd_mem *mem, uint64_t *size)
 {
 	struct amdkfd_process_info *process_info = mem->process_info;
 	unsigned long bo_size = mem->bo->tbo.mem.size;
@@ -1286,9 +1286,11 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
 	struct ttm_validate_buffer *bo_list_entry;
 	unsigned int mapped_to_gpu_memory;
 	int ret;
+	bool is_imported = 0;
 
 	mutex_lock(&mem->lock);
 	mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
+	is_imported = mem->is_imported;
 	mutex_unlock(&mem->lock);
 	/* lock is not needed after this, since mem is unused and will
 	 * be freed anyway
@@ -1340,6 +1342,17 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
 		kfree(mem->bo->tbo.sg);
 	}
 
+	/* Update the size of the BO being freed if it was allocated from
+	 * VRAM and is not imported.
+	 */
+	if (size) {
+		if ((mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM) &&
+		    (!is_imported))
+			*size = bo_size;
+		else
+			*size = 0;
+	}
+
 	/* Free the BO*/
 	amdgpu_bo_unref(&mem->bo);
 	mutex_destroy(&mem->lock);
@@ -1694,6 +1707,7 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd,
 	(*mem)->process_info = avm->process_info;
 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
 	amdgpu_sync_create(&(*mem)->sync);
+	(*mem)->is_imported = true;
 
 	return 0;
 }
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index ff47b1f69b68..cf0017f4d9d5 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -1323,6 +1323,10 @@ static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
 		goto err_free;
 	}
 
+	/* Update the VRAM usage count */
+	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
+		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + args->size);
+
 	mutex_unlock(&p->mutex);
 
 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
@@ -1338,7 +1342,7 @@ static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
 	return 0;
 
 err_free:
-	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem);
+	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem, NULL);
 err_unlock:
 	mutex_unlock(&p->mutex);
 	return err;
@@ -1352,6 +1356,7 @@ static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
 	void *mem;
 	struct kfd_dev *dev;
 	int ret;
+	uint64_t size = 0;
 
 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
 	if (!dev)
@@ -1374,7 +1379,7 @@ static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
 	}
 
 	ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd,
-						(struct kgd_mem *)mem);
+						(struct kgd_mem *)mem, &size);
 
 	/* If freeing the buffer failed, leave the handle in place for
 	 * clean-up during process tear-down.
@@ -1383,6 +1388,8 @@ static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
 		kfd_process_device_remove_obj_handle(
 			pdd, GET_IDR_HANDLE(args->handle));
 
+	WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size);
+
 err_unlock:
 	mutex_unlock(&p->mutex);
 	return ret;
@@ -1727,7 +1734,7 @@ static int kfd_ioctl_import_dmabuf(struct file *filep,
 	return 0;
 
 err_free:
-	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem);
+	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem, NULL);
 err_unlock:
 	mutex_unlock(&p->mutex);
 	return r;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index d48b33449267..cde5e4c7caa1 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -629,6 +629,8 @@ enum kfd_pdd_bound {
 	PDD_BOUND_SUSPENDED,
 };
 
+#define MAX_VRAM_FILENAME_LEN 11
+
 /* Data that is per-process-per device. */
 struct kfd_process_device {
 	/*
@@ -671,6 +673,11 @@ struct kfd_process_device {
 
 	/* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
 	enum kfd_pdd_bound bound;
+
+	/* VRAM usage */
+	uint64_t vram_usage;
+	struct attribute attr_vram;
+	char vram_filename[MAX_VRAM_FILENAME_LEN];
 };
 
 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index 82b4c5a9382a..598296034b43 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -79,18 +79,22 @@ static struct kfd_procfs_tree procfs;
 static ssize_t kfd_procfs_show(struct kobject *kobj, struct attribute *attr,
 			       char *buffer)
 {
-	int val = 0;
-
 	if (strcmp(attr->name, "pasid") == 0) {
 		struct kfd_process *p = container_of(attr, struct kfd_process,
 						     attr_pasid);
-		val = p->pasid;
+
+		return snprintf(buffer, PAGE_SIZE, "%d\n", p->pasid);
+	} else if (strncmp(attr->name, "vram_", 5) == 0) {
+		struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
+							      attr_vram);
+		if (pdd)
+			return snprintf(buffer, PAGE_SIZE, "%llu\n", READ_ONCE(pdd->vram_usage));
 	} else {
 		pr_err("Invalid attribute");
 		return -EINVAL;
 	}
 
-	return snprintf(buffer, PAGE_SIZE, "%d\n", val);
+	return 0;
 }
 
 static void kfd_procfs_kobj_release(struct kobject *kobj)
@@ -206,6 +210,34 @@ int kfd_procfs_add_queue(struct queue *q)
 	return 0;
 }
 
+int kfd_procfs_add_vram_usage(struct kfd_process *p)
+{
+	int ret = 0;
+	struct kfd_process_device *pdd;
+
+	if (!p)
+		return -EINVAL;
+
+	if (!p->kobj)
+		return -EFAULT;
+
+	/* Create proc/<pid>/vram_<gpuid> file for each GPU */
+	list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
+		snprintf(pdd->vram_filename, MAX_VRAM_FILENAME_LEN, "vram_%u",
+			 pdd->dev->id);
+		pdd->attr_vram.name = pdd->vram_filename;
+		pdd->attr_vram.mode = KFD_SYSFS_FILE_MODE;
+		sysfs_attr_init(&pdd->attr_vram);
+		ret = sysfs_create_file(p->kobj, &pdd->attr_vram);
+		if (ret)
+			pr_warn("Creating vram usage for gpu id %d failed",
+				(int)pdd->dev->id);
+	}
+
+	return ret;
+}
+
+
 void kfd_procfs_del_queue(struct queue *q)
 {
 	if (!q)
@@ -248,7 +280,7 @@ static void kfd_process_free_gpuvm(struct kgd_mem *mem,
 	struct kfd_dev *dev = pdd->dev;
 
 	amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(dev->kgd, mem, pdd->vm);
-	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, mem);
+	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, mem, NULL);
 }
 
 /* kfd_process_alloc_gpuvm - Allocate GPU VM for the KFD process
@@ -312,7 +344,7 @@ sync_memory_failed:
 	return err;
 
 err_map_mem:
-	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(kdev->kgd, mem);
+	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(kdev->kgd, mem, NULL);
 err_alloc_mem:
 	*kptr = NULL;
 	return err;
@@ -411,6 +443,11 @@ struct kfd_process *kfd_create_process(struct file *filep)
 							process->kobj);
 		if (!process->kobj_queues)
 			pr_warn("Creating KFD proc/queues folder failed");
+
+		ret = kfd_procfs_add_vram_usage(process);
+		if (ret)
+			pr_warn("Creating vram usage file for pid %d failed",
+				(int)process->lead_thread->pid);
 	}
 out:
 	if (!IS_ERR(process))
@@ -488,7 +525,7 @@ static void kfd_process_device_free_bos(struct kfd_process_device *pdd)
 				peer_pdd->dev->kgd, mem, peer_pdd->vm);
 		}
 
-		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->kgd, mem);
+		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->kgd, mem, NULL);
 		kfd_process_device_remove_obj_handle(pdd, id);
 	}
 }
@@ -551,6 +588,7 @@ static void kfd_process_wq_release(struct work_struct *work)
 {
 	struct kfd_process *p = container_of(work, struct kfd_process,
 					     release_work);
+	struct kfd_process_device *pdd;
 
 	/* Remove the procfs files */
 	if (p->kobj) {
@@ -558,6 +596,10 @@ static void kfd_process_wq_release(struct work_struct *work)
 		kobject_del(p->kobj_queues);
 		kobject_put(p->kobj_queues);
 		p->kobj_queues = NULL;
+
+		list_for_each_entry(pdd, &p->per_device_data, per_device_list)
+			sysfs_remove_file(p->kobj, &pdd->attr_vram);
+
 		kobject_del(p->kobj);
 		kobject_put(p->kobj);
 		p->kobj = NULL;
@@ -863,6 +905,7 @@ struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
 	pdd->bound = PDD_UNBOUND;
 	pdd->already_dequeued = false;
 	pdd->runtime_inuse = false;
+	pdd->vram_usage = 0;
 	list_add(&pdd->per_device_list, &p->per_device_data);
 
 	/* Init idr used for memory handle translation */