Re: [PATCH v3 16/18] xen/pvcalls: implement read

[Juergen Gross <jgross@suse.com>]

On 02/06/17 21:31, Stefano Stabellini wrote:
> When an active socket has data available, increment the io and read
> counters, and schedule the ioworker.
> 
> Implement the read function by reading from the socket, writing the data
> to the data ring.
> 
> Set in_error on error.
> 
> Signed-off-by: Stefano Stabellini <stefano@aporeto.com>
> CC: boris.ostrovsky@oracle.com
> CC: jgross@suse.com
> ---
>  drivers/xen/pvcalls-back.c | 85 ++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 85 insertions(+)
> 
> diff --git a/drivers/xen/pvcalls-back.c b/drivers/xen/pvcalls-back.c
> index 0283d49..e7d2b85 100644
> --- a/drivers/xen/pvcalls-back.c
> +++ b/drivers/xen/pvcalls-back.c
> @@ -101,6 +101,81 @@ static int pvcalls_back_release_active(struct xenbus_device *dev,
>  
>  static void pvcalls_conn_back_read(unsigned long opaque)
>  {
> +	struct sock_mapping *map = (struct sock_mapping *)opaque;
> +	struct msghdr msg;
> +	struct kvec vec[2];
> +	RING_IDX cons, prod, size, wanted, array_size, masked_prod, masked_cons;
> +	int32_t error;
> +	struct pvcalls_data_intf *intf = map->ring;
> +	struct pvcalls_data *data = &map->data;
> +	unsigned long flags;
> +	int ret;
> +
> +	array_size = XEN_FLEX_RING_SIZE(map->ring_order);
> +	cons = intf->in_cons;
> +	prod = intf->in_prod;
> +	error = intf->in_error;
> +	/* read the indexes first, then deal with the data */
> +	virt_mb();
> +
> +	if (error)
> +		return;
> +
> +	size = pvcalls_queued(prod, cons, array_size);
> +	if (size >= array_size)
> +		return;
> +	spin_lock_irqsave(&map->sock->sk->sk_receive_queue.lock, flags);
> +	if (skb_queue_empty(&map->sock->sk->sk_receive_queue)) {
> +		atomic_set(&map->read, 0);
> +		spin_unlock_irqrestore(&map->sock->sk->sk_receive_queue.lock,
> +				flags);
> +		return;
> +	}
> +	spin_unlock_irqrestore(&map->sock->sk->sk_receive_queue.lock, flags);
> +	wanted = array_size - size;
> +	masked_prod = pvcalls_mask(prod, array_size);
> +	masked_cons = pvcalls_mask(cons, array_size);
> +
> +	memset(&msg, 0, sizeof(msg));
> +	msg.msg_iter.type = ITER_KVEC|WRITE;
> +	msg.msg_iter.count = wanted;
> +	if (masked_prod < masked_cons) {
> +		vec[0].iov_base = data->in + masked_prod;
> +		vec[0].iov_len = wanted;
> +		msg.msg_iter.kvec = vec;
> +		msg.msg_iter.nr_segs = 1;
> +	} else {
> +		vec[0].iov_base = data->in + masked_prod;
> +		vec[0].iov_len = array_size - masked_prod;
> +		vec[1].iov_base = data->in;
> +		vec[1].iov_len = wanted - vec[0].iov_len;
> +		msg.msg_iter.kvec = vec;
> +		msg.msg_iter.nr_segs = 2;
> +	}
> +
> +	atomic_set(&map->read, 0);
> +	ret = inet_recvmsg(map->sock, &msg, wanted, MSG_DONTWAIT);
> +	WARN_ON(ret > 0 && ret > wanted);

wanted is always > 0, so you can omit the ret > 0 test.


Juergen

> +	if (ret == -EAGAIN) /* shouldn't happen */
> +		return;
> +	if (!ret)
> +		ret = -ENOTCONN;
> +	spin_lock_irqsave(&map->sock->sk->sk_receive_queue.lock, flags);
> +	if (ret > 0 && !skb_queue_empty(&map->sock->sk->sk_receive_queue))
> +		atomic_inc(&map->read);
> +	spin_unlock_irqrestore(&map->sock->sk->sk_receive_queue.lock, flags);
> +
> +	/* write the data, then modify the indexes */
> +	virt_wmb();
> +	if (ret < 0)
> +		intf->in_error = ret;
> +	else
> +		intf->in_prod = prod + ret;
> +	/* update the indexes, then notify the other end */
> +	virt_wmb();
> +	notify_remote_via_irq(map->irq);
> +
> +	return;
>  }
>  
>  static int pvcalls_conn_back_write(struct sock_mapping *map)
> @@ -173,6 +248,16 @@ static void pvcalls_sk_state_change(struct sock *sock)
>  
>  static void pvcalls_sk_data_ready(struct sock *sock)
>  {
> +	struct sock_mapping *map = sock->sk_user_data;
> +	struct pvcalls_ioworker *iow;
> +
> +	if (map == NULL)
> +		return;
> +
> +	iow = &map->ioworker;
> +	atomic_inc(&map->read);
> +	atomic_inc(&map->io);
> +	queue_work_on(iow->cpu, iow->wq, &iow->register_work);
>  }
>  
>  static struct sock_mapping *pvcalls_new_active_socket(
>