Split the function vhost_worker_create to support both task and kthread

Added back the previous old function vhost_worker_create and rename it to
vhost_worker_create_khtread to support the khtread.

The new vhost_worker_create will be selected which to use based on the
value of the parameter.

the old function vhost_worker_create was change to support task in
commit 6e890c5d5021 ("vhost: use vhost_tasks for worker threads")
also changed in
commit 1cdaafa1b8b4 ("vhost: replace single worker pointer with xarray")
commit c011bb669ddc ("vhost: dynamically allocate vhost_worker")

Signed-off-by: Cindy Lu <lulu@redhat.com>
---
 drivers/vhost/vhost.c | 55 ++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 54 insertions(+), 1 deletion(-)

diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index eb30da658bfe..08c9e77916ca 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c
@@ -808,7 +808,8 @@ static void vhost_workers_free(struct vhost_dev *dev)
 	else
 		vhost_workers_free_kthread(dev);
 }
-static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev)
+
+static struct vhost_worker *vhost_worker_create_task(struct vhost_dev *dev)
 {
 	struct vhost_worker *worker;
 	struct vhost_task *vtsk;
@@ -849,6 +850,50 @@ static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev)
 	return NULL;
 }
 
+static struct vhost_worker *vhost_worker_create_kthread(struct vhost_dev *dev)
+{
+	struct vhost_worker *worker;
+	struct task_struct *task;
+	int ret;
+	u32 id;
+
+	worker = kzalloc(sizeof(*worker), GFP_KERNEL_ACCOUNT);
+	if (!worker)
+		return NULL;
+
+	worker->dev = dev;
+	worker->kcov_handle = kcov_common_handle();
+
+	mutex_init(&worker->mutex);
+	init_llist_head(&worker->work_list);
+
+	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;
+	wake_up_process(task); /* avoid contributing to loadavg */
+	ret = xa_alloc(&dev->worker_xa, &id, worker, xa_limit_32b, GFP_KERNEL);
+	if (ret < 0)
+		goto stop_worker;
+	worker->id = id;
+
+	ret = vhost_attach_cgroups(dev);
+	if (ret)
+		goto stop_worker;
+
+	return worker;
+
+stop_worker:
+	kthread_stop(worker->task);
+free_worker:
+	kfree(worker);
+	return NULL;
+}
+
 /* Caller must have device mutex */
 static void __vhost_vq_attach_worker(struct vhost_virtqueue *vq,
 				     struct vhost_worker *worker)
@@ -937,6 +982,14 @@ static int vhost_vq_attach_worker(struct vhost_virtqueue *vq,
 	return 0;
 }
 
+static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev)
+{
+	if (enforce_inherit_owner)
+		return vhost_worker_create_task(dev);
+	else
+		return vhost_worker_create_kthread(dev);
+}
+
 /* Caller must have device mutex */
 static int vhost_new_worker(struct vhost_dev *dev,
 			    struct vhost_worker_state *info)
-- 
2.45.0

On 10/3/24 8:58 PM, Cindy Lu wrote:
> +static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev)
> +{
> +	if (enforce_inherit_owner)
> +		return vhost_worker_create_task(dev);
> +	else
> +		return vhost_worker_create_kthread(dev);
> +}

The reason we are in this whole mess is because for vhost-scsi
we wanted to create multiple worker threads as the single thread
became a bottleneck. We couldn't use kthreads because they were
not accounted for under thread that created it and we could
exceed the NPROC_LIMIT.

So with the above code, we could hit that problem again if userspace
did the VHOST_NEW_WORKER ioctl and enforce_inherit_owner=false.

You would want a check for enforce_inherit_owner in vhost_worker_ioctl.

On Tue, 15 Oct 2024 at 05:02, Mike Christie <michael.christie@oracle.com> wrote:
>
> On 10/3/24 8:58 PM, Cindy Lu wrote:
> > +static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev)
> > +{
> > +     if (enforce_inherit_owner)
> > +             return vhost_worker_create_task(dev);
> > +     else
> > +             return vhost_worker_create_kthread(dev);
> > +}
>
> The reason we are in this whole mess is because for vhost-scsi
> we wanted to create multiple worker threads as the single thread
> became a bottleneck. We couldn't use kthreads because they were
> not accounted for under thread that created it and we could
> exceed the NPROC_LIMIT.
>
> So with the above code, we could hit that problem again if userspace
> did the VHOST_NEW_WORKER ioctl and enforce_inherit_owner=false.
>
> You would want a check for enforce_inherit_owner in vhost_worker_ioctl.
>
sure, will add this check
Thanks
cindy