On 11/5/24 1:25 AM, Cindy Lu wrote:

 static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev)
 {
 	struct vhost_worker *worker;
-	struct vhost_task *vtsk;
+	struct vhost_task *vtsk = NULL;
+	struct task_struct *task = NULL;
 	char name[TASK_COMM_LEN];
 	int ret;
 	u32 id;
 
+	/* Allocate resources for the worker */
 	worker = kzalloc(sizeof(*worker), GFP_KERNEL_ACCOUNT);
 	if (!worker)
 		return NULL;
 
+	worker->fn = kzalloc(sizeof(struct vhost_task_fn), GFP_KERNEL_ACCOUNT);
+	if (!worker->fn) {
+		kfree(worker);
+		return NULL;
+	}

Why dynamically allocate this?

You could probably even just kill the vhost_task_fn struct since we just
have to the 2 callouts.

+
 	worker->dev = dev;
 	snprintf(name, sizeof(name), "vhost-%d", current->pid);
 
-	vtsk = vhost_task_create(vhost_run_work_list, vhost_worker_killed,
-				 worker, name);
-	if (!vtsk)
-		goto free_worker;
-
 	mutex_init(&worker->mutex);
 	init_llist_head(&worker->work_list);
 	worker->kcov_handle = kcov_common_handle();
-	worker->vtsk = vtsk;
 
-	vhost_task_start(vtsk);
+	if (dev->inherit_owner) {
+		/* Create and start a vhost task */

Maybe instead of this comment and the one below write something about
what inherit_owner means. We can see from the code we are creating a
vhost/kthread, but it's not really obvious why. Something like:

/*
 * If inherit_owner is true we use vhost_tasks to create
 * the worker so all settings/limits like cgroups, NPROC,
 * scheduler, etc are inherited from the owner. If false,
 * we use kthreads and only attach to the same cgroups
 * as the owner for compat with older kernels.
 */

+		vtsk = vhost_task_create(vhost_run_work_list,
+					 vhost_worker_killed, worker, name);
+		if (!vtsk)
+			goto free_worker;
+
+		worker->vtsk = vtsk;
+		worker->fn->wakeup = vhost_task_wakeup_fn;
+		worker->fn->stop = vhost_task_stop_fn;
+
+		vhost_task_start(vtsk);
+	} else {
+		/* Create and start a kernel thread */
+		task = kthread_create(vhost_run_work_kthread_list, worker,
+				      "vhost-%d", current->pid);
+		if (IS_ERR(task)) {
+			ret = PTR_ERR(task);
+			goto free_worker;
+		}
+		worker->task = task;
+		worker->fn->wakeup = vhost_kthread_wakeup_fn;
+		worker->fn->stop = vhost_kthread_stop_fn;
+
+		wake_up_process(task);
+		/* Attach to the vhost cgroup */

You don't need this comment do you? The function name tells us the same
info.

+		ret = vhost_attach_cgroups(dev);

I don't think this works. Patch 3/9 did:

+	xa_for_each(&dev->worker_xa, i, worker) {
+		ret = vhost_worker_cgroups_kthread(worker);

but we don't add the worker to the xa until below.

You also want to just call vhost_worker_cgroups_kthread above, because
you only want to add the one task and not loop over all of them.

I would then also maybe rename vhost_worker_cgroups_kthread to something
like vhost_attach_task_to_cgroups.

+		if (ret)
+			goto stop_worker;
+	}
 
 	ret = xa_alloc(&dev->worker_xa, &id, worker, xa_limit_32b, GFP_KERNEL);
 	if (ret < 0)
 		goto stop_worker;
 	worker->id = id;
-
 	return worker;
-
 stop_worker:
-	vhost_task_stop(vtsk);
+	worker->fn->stop(dev->inherit_owner ? (void *)vtsk : (void *)task);

I don't think you need to cast since the function takes a void pointer.
Same comment for the other patches like 6/9 where you are calling the
callout and casting.