Re: [PATCH v3 11/18] xen/pvcalls: implement accept command

[Juergen Gross <jgross@suse.com>]

On 02/06/17 21:31, Stefano Stabellini wrote:
> Implement the accept command by calling inet_accept. To avoid blocking
> in the kernel, call inet_accept(O_NONBLOCK) from a workqueue, which get
> scheduled on sk_data_ready (for a passive socket, it means that there
> are connections to accept).
> 
> Use the reqcopy field to store the request. Accept the new socket from
> the delayed work function, create a new sock_mapping for it, map
> the indexes page and data ring, and reply to the other end. Allocate an
> ioworker for the socket.
> 
> Only support one outstanding blocking accept request for every socket at
> any time.
> 
> Add a field to sock_mapping to remember the passive socket from which an
> active socket was created.
> 
> Signed-off-by: Stefano Stabellini <stefano@aporeto.com>
> CC: boris.ostrovsky@oracle.com
> CC: jgross@suse.com
> ---
>  drivers/xen/pvcalls-back.c | 109 ++++++++++++++++++++++++++++++++++++++++++++-
>  1 file changed, 108 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/xen/pvcalls-back.c b/drivers/xen/pvcalls-back.c
> index a75586e..f1173f4 100644
> --- a/drivers/xen/pvcalls-back.c
> +++ b/drivers/xen/pvcalls-back.c
> @@ -65,6 +65,7 @@ struct pvcalls_ioworker {
>  struct sock_mapping {
>  	struct list_head list;
>  	struct pvcalls_fedata *priv;
> +	struct sockpass_mapping *sockpass;
>  	struct socket *sock;
>  	uint64_t id;
>  	grant_ref_t ref;
> @@ -275,10 +276,79 @@ static int pvcalls_back_release(struct xenbus_device *dev,
>  
>  static void __pvcalls_back_accept(struct work_struct *work)
>  {
> +	struct sockpass_mapping *mappass = container_of(
> +		work, struct sockpass_mapping, register_work);
> +	struct sock_mapping *map;
> +	struct pvcalls_ioworker *iow;
> +	struct pvcalls_fedata *priv;
> +	struct socket *sock;
> +	struct xen_pvcalls_response *rsp;
> +	struct xen_pvcalls_request *req;
> +	int notify;
> +	int ret = -EINVAL;
> +	unsigned long flags;
> +
> +	priv = mappass->priv;
> +	/* We only need to check the value of "cmd" atomically on read. */
> +	spin_lock_irqsave(&mappass->copy_lock, flags);
> +	req = &mappass->reqcopy;
> +	if (req->cmd != PVCALLS_ACCEPT) {
> +		spin_unlock_irqrestore(&mappass->copy_lock, flags);
> +		return;
> +	}
> +	spin_unlock_irqrestore(&mappass->copy_lock, flags);

What about:
	req = &mappass->reqcopy;
	if (ACCESS_ONCE(req->cmd) != PVCALLS_ACCEPT)
		return;

I can't see the need for taking a lock here.

> +
> +	sock = sock_alloc();
> +	if (sock == NULL)
> +		goto out_error;
> +	sock->type = mappass->sock->type;
> +	sock->ops = mappass->sock->ops;
> +
> +	ret = inet_accept(mappass->sock, sock, O_NONBLOCK, true);
> +	if (ret == -EAGAIN) {
> +		sock_release(sock);
> +		goto out_error;
> +	}
> +
> +	map = pvcalls_new_active_socket(priv,
> +					req->u.accept.id_new,
> +					req->u.accept.ref,
> +					req->u.accept.evtchn,
> +					sock);
> +	if (!map) {
> +		sock_release(sock);
> +		goto out_error;
> +	}
> +
> +	map->sockpass = mappass;
> +	iow = &map->ioworker;
> +	atomic_inc(&map->read);
> +	atomic_inc(&map->io);
> +	queue_work_on(iow->cpu, iow->wq, &iow->register_work);
> +
> +out_error:
> +	rsp = RING_GET_RESPONSE(&priv->ring, priv->ring.rsp_prod_pvt++);
> +	rsp->req_id = req->req_id;
> +	rsp->cmd = req->cmd;
> +	rsp->u.accept.id = req->u.accept.id;
> +	rsp->ret = ret;
> +	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&priv->ring, notify);
> +	if (notify)
> +		notify_remote_via_irq(priv->irq);
> +
> +	spin_lock_irqsave(&mappass->copy_lock, flags);
> +	mappass->reqcopy.cmd = 0;
> +	spin_unlock_irqrestore(&mappass->copy_lock, flags);

ACCESS_ONCE(mappass->reqcopy.cmd) = 0;

>  }
>  
>  static void pvcalls_pass_sk_data_ready(struct sock *sock)
>  {
> +	struct sockpass_mapping *mappass = sock->sk_user_data;
> +
> +	if (mappass == NULL)
> +		return;
> +
> +	queue_work(mappass->wq, &mappass->register_work);
>  }
>  
>  static int pvcalls_back_bind(struct xenbus_device *dev,
> @@ -380,7 +450,44 @@ static int pvcalls_back_listen(struct xenbus_device *dev,
>  static int pvcalls_back_accept(struct xenbus_device *dev,
>  			       struct xen_pvcalls_request *req)
>  {
> -	return 0;
> +	struct pvcalls_fedata *priv;
> +	struct sockpass_mapping *mappass;
> +	int ret = -EINVAL;
> +	struct xen_pvcalls_response *rsp;
> +	unsigned long flags;
> +
> +	priv = dev_get_drvdata(&dev->dev);
> +
> +	mappass = radix_tree_lookup(&priv->socketpass_mappings,
> +		req->u.accept.id);
> +	if (mappass == NULL)
> +		goto out_error;
> +
> +	/* 
> +	 * Limitation of the current implementation: only support one
> +	 * concurrent accept or poll call on one socket.
> +	 */
> +	spin_lock_irqsave(&mappass->copy_lock, flags);
> +	if (mappass->reqcopy.cmd != 0) {
> +		spin_unlock_irqrestore(&mappass->copy_lock, flags);
> +		ret = -EINTR;
> +		goto out_error;
> +	}
> +
> +	mappass->reqcopy = *req;

This time you need the lock, however you should use:

ACCESS_ONCE(mappass->reqcopy) = *req;

> +	spin_unlock_irqrestore(&mappass->copy_lock, flags);
> +	queue_work(mappass->wq, &mappass->register_work);
> +
> +	/* Tell the caller we don't need to send back a notification yet */
> +	return -1;
> +
> +out_error:
> +	rsp = RING_GET_RESPONSE(&priv->ring, priv->ring.rsp_prod_pvt++);
> +	rsp->req_id = req->req_id;
> +	rsp->cmd = req->cmd;
> +	rsp->u.accept.id = req->u.accept.id;
> +	rsp->ret = ret;
> +	return ret;

return 0?


Juergen

>  }
>  
>  static int pvcalls_back_poll(struct xenbus_device *dev,
>