Re: [PATCH v9 2/4] vhost: Reintroduce kthread mode support in vhost

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, Apr 21, 2025 at 10:44:08AM +0800, Cindy Lu wrote:
> This patch reintroduces kthread mode support in vhost,
> It also introduces struct vhost_worker_ops to abstract
> worker create/stop/wakeup operations.
> 
> * Bring back the original vhost_worker() implementation,
>   and renamed to vhost_run_work_kthread_list().
> 
> * Add cgroup support for the kthread
> 
> * Introduce struct vhost_worker_ops:
>   - Encapsulates create / stop / wake‑up callbacks.
>   - vhost_worker_create() selects the proper ops according to
>     inherit_owner.
> 
> This partially reverts or improves upon:
> commit 6e890c5d5021 ("vhost: use vhost_tasks for worker threads")
> commit 1cdaafa1b8b4 ("vhost: replace single worker pointer with xarray")
> 
> Signed-off-by: Cindy Lu <lulu@redhat.com>
> ---
>  drivers/vhost/vhost.c | 188 ++++++++++++++++++++++++++++++++++++++----
>  drivers/vhost/vhost.h |  12 +++
>  2 files changed, 182 insertions(+), 18 deletions(-)
> 
> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> index 250dc43f1786..be97028a8baf 100644
> --- a/drivers/vhost/vhost.c
> +++ b/drivers/vhost/vhost.c
> @@ -22,6 +22,7 @@
>  #include <linux/slab.h>
>  #include <linux/vmalloc.h>
>  #include <linux/kthread.h>
> +#include <linux/cgroup.h>
>  #include <linux/module.h>
>  #include <linux/sort.h>
>  #include <linux/sched/mm.h>
> @@ -242,7 +243,7 @@ static void vhost_worker_queue(struct vhost_worker *worker,
>  		 * test_and_set_bit() implies a memory barrier.
>  		 */
>  		llist_add(&work->node, &worker->work_list);
> -		vhost_task_wake(worker->vtsk);
> +		worker->ops->wakeup(worker);
>  	}
>  }
>  
> @@ -388,6 +389,44 @@ static void vhost_vq_reset(struct vhost_dev *dev,
>  	__vhost_vq_meta_reset(vq);
>  }
>  
> +static int vhost_run_work_kthread_list(void *data)
> +{
> +	struct vhost_worker *worker = data;
> +	struct vhost_work *work, *work_next;
> +	struct vhost_dev *dev = worker->dev;
> +	struct llist_node *node;
> +
> +	kthread_use_mm(dev->mm);
> +
> +	for (;;) {
> +		/* mb paired w/ kthread_stop */
> +		set_current_state(TASK_INTERRUPTIBLE);
> +
> +		if (kthread_should_stop()) {
> +			__set_current_state(TASK_RUNNING);
> +			break;
> +		}
> +		node = llist_del_all(&worker->work_list);
> +		if (!node)
> +			schedule();
> +
> +		node = llist_reverse_order(node);
> +		/* make sure flag is seen after deletion */
> +		smp_wmb();
> +		llist_for_each_entry_safe(work, work_next, node, node) {
> +			clear_bit(VHOST_WORK_QUEUED, &work->flags);
> +			__set_current_state(TASK_RUNNING);
> +			kcov_remote_start_common(worker->kcov_handle);
> +			work->fn(work);
> +			kcov_remote_stop();
> +			cond_resched();
> +		}
> +	}
> +	kthread_unuse_mm(dev->mm);
> +
> +	return 0;
> +}
> +
>  static bool vhost_run_work_list(void *data)
>  {
>  	struct vhost_worker *worker = data;
> @@ -582,6 +621,46 @@ long vhost_dev_check_owner(struct vhost_dev *dev)
>  }
>  EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
>  
> +struct vhost_attach_cgroups_struct {
> +	struct vhost_work work;
> +	struct task_struct *owner;
> +	int ret;
> +};
> +
> +static void vhost_attach_cgroups_work(struct vhost_work *work)
> +{
> +	struct vhost_attach_cgroups_struct *s;
> +
> +	s = container_of(work, struct vhost_attach_cgroups_struct, work);
> +	s->ret = cgroup_attach_task_all(s->owner, current);
> +}
> +
> +static int vhost_attach_task_to_cgroups(struct vhost_worker *worker)
> +{
> +	struct vhost_attach_cgroups_struct attach;
> +	int saved_cnt;
> +
> +	attach.owner = current;
> +
> +	vhost_work_init(&attach.work, vhost_attach_cgroups_work);
> +	vhost_worker_queue(worker, &attach.work);
> +
> +	mutex_lock(&worker->mutex);
> +
> +	/*
> +	 * Bypass attachment_cnt check in __vhost_worker_flush:
> +	 * Temporarily change it to INT_MAX to bypass the check
> +	 */
> +	saved_cnt = worker->attachment_cnt;
> +	worker->attachment_cnt = INT_MAX;
> +	__vhost_worker_flush(worker);
> +	worker->attachment_cnt = saved_cnt;


You mean this one?
        if (!worker->attachment_cnt || worker->killed)
                return;


Just introduce a variant of __vhost_worker_flush that
skips this check.

E.g.

Rename __vhost_worker_flush -> _vhost_worker_flush.

then rework:

static void _vhost_worker_flush(struct vhost_worker *worker)
{
        struct vhost_flush_struct flush; 
                                  
        if (!worker->attachment_cnt || worker->killed)
                return;

	__vhost_worker_flush(worker);
}





> +
> +	mutex_unlock(&worker->mutex);
> +
> +	return attach.ret;
> +}
> +
>  /* Caller should have device mutex */
>  bool vhost_dev_has_owner(struct vhost_dev *dev)
>  {
> @@ -627,7 +706,7 @@ static void vhost_worker_destroy(struct vhost_dev *dev,
>  
>  	WARN_ON(!llist_empty(&worker->work_list));
>  	xa_erase(&dev->worker_xa, worker->id);
> -	vhost_task_stop(worker->vtsk);
> +	worker->ops->stop(worker);
>  	kfree(worker);
>  }
>  
> @@ -650,42 +729,115 @@ static void vhost_workers_free(struct vhost_dev *dev)
>  	xa_destroy(&dev->worker_xa);
>  }
>  
> +static void vhost_task_wakeup(struct vhost_worker *worker)
> +{
> +	return vhost_task_wake(worker->vtsk);
> +}
> +
> +static void vhost_kthread_wakeup(struct vhost_worker *worker)
> +{
> +	wake_up_process(worker->kthread_task);
> +}
> +
> +static void vhost_task_do_stop(struct vhost_worker *worker)
> +{
> +	return vhost_task_stop(worker->vtsk);
> +}
> +
> +static void vhost_kthread_do_stop(struct vhost_worker *worker)
> +{
> +	kthread_stop(worker->kthread_task);
> +}
> +
> +static int vhost_task_worker_create(struct vhost_worker *worker,
> +				    struct vhost_dev *dev, const char *name)
> +{
> +	struct vhost_task *vtsk;
> +	u32 id;
> +	int ret;
> +
> +	vtsk = vhost_task_create(vhost_run_work_list, vhost_worker_killed,
> +				 worker, name);
> +	if (IS_ERR(vtsk))
> +		return PTR_ERR(vtsk);
> +
> +	worker->vtsk = vtsk;
> +	vhost_task_start(vtsk);
> +	ret = xa_alloc(&dev->worker_xa, &id, worker, xa_limit_32b, GFP_KERNEL);
> +	if (ret < 0) {
> +		vhost_task_do_stop(worker);
> +		return ret;
> +	}
> +	worker->id = id;
> +	return 0;
> +}
> +
> +static int vhost_kthread_worker_create(struct vhost_worker *worker,
> +				       struct vhost_dev *dev, const char *name)
> +{
> +	struct task_struct *task;
> +	u32 id;
> +	int ret;
> +
> +	task = kthread_create(vhost_run_work_kthread_list, worker, "%s", name);
> +	if (IS_ERR(task))
> +		return PTR_ERR(task);
> +
> +	worker->kthread_task = task;
> +	wake_up_process(task);
> +	ret = xa_alloc(&dev->worker_xa, &id, worker, xa_limit_32b, GFP_KERNEL);
> +	if (ret < 0)
> +		goto stop_worker;
> +
> +	ret = vhost_attach_task_to_cgroups(worker);
> +	if (ret)
> +		goto stop_worker;
> +
> +	worker->id = id;
> +	return 0;
> +
> +stop_worker:
> +	vhost_kthread_do_stop(worker);
> +	return ret;
> +}
> +
> +static const struct vhost_worker_ops kthread_ops = {
> +	.create = vhost_kthread_worker_create,
> +	.stop = vhost_kthread_do_stop,
> +	.wakeup = vhost_kthread_wakeup,
> +};
> +
> +static const struct vhost_worker_ops vhost_task_ops = {
> +	.create = vhost_task_worker_create,
> +	.stop = vhost_task_do_stop,
> +	.wakeup = vhost_task_wakeup,
> +};
> +
>  static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev)
>  {
>  	struct vhost_worker *worker;
> -	struct vhost_task *vtsk;
>  	char name[TASK_COMM_LEN];
>  	int ret;
> -	u32 id;
> +	const struct vhost_worker_ops *ops =
> +		dev->inherit_owner ? &vhost_task_ops : &kthread_ops;
>  
>  	worker = kzalloc(sizeof(*worker), GFP_KERNEL_ACCOUNT);
>  	if (!worker)
>  		return NULL;
>  
>  	worker->dev = dev;
> +	worker->ops = ops;
>  	snprintf(name, sizeof(name), "vhost-%d", current->pid);
>  
> -	vtsk = vhost_task_create(vhost_run_work_list, vhost_worker_killed,
> -				 worker, name);
> -	if (IS_ERR(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);
> -
> -	ret = xa_alloc(&dev->worker_xa, &id, worker, xa_limit_32b, GFP_KERNEL);
> +	ret = ops->create(worker, dev, name);
>  	if (ret < 0)
> -		goto stop_worker;
> -	worker->id = id;
> +		goto free_worker;
>  
>  	return worker;
>  
> -stop_worker:
> -	vhost_task_stop(vtsk);
>  free_worker:
>  	kfree(worker);
>  	return NULL;
> diff --git a/drivers/vhost/vhost.h b/drivers/vhost/vhost.h
> index 19bb94922a0e..af4b2f7d3b91 100644
> --- a/drivers/vhost/vhost.h
> +++ b/drivers/vhost/vhost.h
> @@ -26,7 +26,18 @@ struct vhost_work {
>  	unsigned long		flags;
>  };
>  
> +struct vhost_worker;
> +struct vhost_dev;
> +
> +struct vhost_worker_ops {
> +	int (*create)(struct vhost_worker *worker, struct vhost_dev *dev,
> +		      const char *name);
> +	void (*stop)(struct vhost_worker *worker);
> +	void (*wakeup)(struct vhost_worker *worker);
> +};
> +
>  struct vhost_worker {
> +	struct task_struct *kthread_task;
>  	struct vhost_task	*vtsk;
>  	struct vhost_dev	*dev;
>  	/* Used to serialize device wide flushing with worker swapping. */
> @@ -36,6 +47,7 @@ struct vhost_worker {
>  	u32			id;
>  	int			attachment_cnt;
>  	bool			killed;
> +	const struct vhost_worker_ops *ops;
>  };
>  
>  /* Poll a file (eventfd or socket) */
> -- 
> 2.45.0