Re: [PATCH v3 09/13] xen/pvcalls: implement sendmsg

[Boris Ostrovsky <boris.ostrovsky@oracle.com>]

On 07/31/2017 06:57 PM, Stefano Stabellini wrote:
> Send data to an active socket by copying data to the "out" ring. Take
> the active socket out_mutex so that only one function can access the
> ring at any given time.
> 
> If not enough room is available on the ring, rather than returning
> immediately or sleep-waiting, spin for up to 5000 cycles. This small
> optimization turns out to improve performance significantly.
> 
> Signed-off-by: Stefano Stabellini <stefano@aporeto.com>
> CC: boris.ostrovsky@oracle.com
> CC: jgross@suse.com
> ---
>   drivers/xen/pvcalls-front.c | 109 ++++++++++++++++++++++++++++++++++++++++++++
>   drivers/xen/pvcalls-front.h |   3 ++
>   2 files changed, 112 insertions(+)
> 
> diff --git a/drivers/xen/pvcalls-front.c b/drivers/xen/pvcalls-front.c
> index f83b910..369acde 100644
> --- a/drivers/xen/pvcalls-front.c
> +++ b/drivers/xen/pvcalls-front.c
> @@ -29,6 +29,7 @@
>   #define PVCALLS_INVALID_ID UINT_MAX
>   #define PVCALLS_RING_ORDER XENBUS_MAX_RING_GRANT_ORDER
>   #define PVCALLS_NR_REQ_PER_RING __CONST_RING_SIZE(xen_pvcalls, XEN_PAGE_SIZE)
> +#define PVCALLS_FRONT_MAX_SPIN 5000
>   
>   struct pvcalls_bedata {
>   	struct xen_pvcalls_front_ring ring;
> @@ -88,6 +89,22 @@ static inline int get_request(struct pvcalls_bedata *bedata, int *req_id)
>   	return 0;
>   }
>   
> +static int pvcalls_front_write_todo(struct sock_mapping *map)
> +{
> +	struct pvcalls_data_intf *intf = map->active.ring;
> +	RING_IDX cons, prod, size = XEN_FLEX_RING_SIZE(intf->ring_order);
> +	int32_t error;
> +
> +	cons = intf->out_cons;
> +	prod = intf->out_prod;
> +	error = intf->out_error;
> +	if (error == -ENOTCONN)
> +		return 0;
> +	if (error != 0)
> +		return error;
> +	return size - pvcalls_queued(prod, cons, size);

Do you ever look at actual return value except whether it's zero or not? 
Both here and in the poll patch you look for !=0 and never check for an 
error.

> +}
> +
>   static irqreturn_t pvcalls_front_event_handler(int irq, void *dev_id)
>   {
>   	struct xenbus_device *dev = dev_id;
> @@ -325,6 +342,98 @@ int pvcalls_front_connect(struct socket *sock, struct sockaddr *addr,
>   	return ret;
>   }
>   
> +static int __write_ring(struct pvcalls_data_intf *intf,
> +			struct pvcalls_data *data,
> +			struct iov_iter *msg_iter,
> +			size_t len)
> +{
> +	RING_IDX cons, prod, size, masked_prod, masked_cons;
> +	RING_IDX array_size = XEN_FLEX_RING_SIZE(intf->ring_order);
> +	int32_t error;
> +
> +	cons = intf->out_cons;
> +	prod = intf->out_prod;
> +	error = intf->out_error;
> +	/* read indexes before continuing */
> +	virt_mb();
> +
> +	if (error < 0)
> +		return error;

This can be done before the barrier. (In fact, this can be done first 
thing, before cons and prod are read).

> +
> +	size = pvcalls_queued(prod, cons, array_size);
> +	if (size >= array_size)
> +		return 0;
> +	if (len > array_size - size)
> +		len = array_size - size;
> +
> +	masked_prod = pvcalls_mask(prod, array_size);
> +	masked_cons = pvcalls_mask(cons, array_size);
> +
> +	if (masked_prod < masked_cons) {
> +		copy_from_iter(data->out + masked_prod, len, msg_iter);
> +	} else {
> +		if (len > array_size - masked_prod) {
> +			copy_from_iter(data->out + masked_prod,
> +				       array_size - masked_prod, msg_iter);
> +			copy_from_iter(data->out,
> +				       len - (array_size - masked_prod),
> +				       msg_iter);
> +		} else {
> +			copy_from_iter(data->out + masked_prod, len, msg_iter);
> +		}
> +	}
> +	/* write to ring before updating pointer */
> +	virt_wmb();
> +	intf->out_prod += len;
> +
> +	return len;

You are returning size_t. I actually was going to ask in one of the 
previous patches whether using int for sizes was correct. Unfortunately 
I can't remember which struct/function I was looking at ;-(

Of course, you are also possibly returning a (negative) error here.

> +}
> +
> +int pvcalls_front_sendmsg(struct socket *sock, struct msghdr *msg,
> +			  size_t len)
> +{
> +	struct pvcalls_bedata *bedata;
> +	struct sock_mapping *map;
> +	int sent = 0, tot_sent = 0;

'sent' doesn't need to be initialized.

> +	int count = 0, flags;
> +
> +	if (!pvcalls_front_dev)
> +		return -ENOTCONN;
> +	bedata = dev_get_drvdata(&pvcalls_front_dev->dev);
> +
> +	map = (struct sock_mapping *) READ_ONCE(sock->sk->sk_send_head);
> +	if (!map)
> +		return -ENOTSOCK;

IIRC the error value for sk_send_head being zero is inconsistent across 
patches.

> +
> +	flags = msg->msg_flags;
> +	if (flags & (MSG_CONFIRM|MSG_DONTROUTE|MSG_EOR|MSG_OOB))
> +		return -EOPNOTSUPP;
> +
> +	mutex_lock(&map->active.out_mutex);
> +	if ((flags & MSG_DONTWAIT) && !pvcalls_front_write_todo(map)) {
> +		mutex_unlock(&map->active.out_mutex);
> +		return -EAGAIN;
> +	}
> +
> +again:
> +	count++;
> +	sent = __write_ring(map->active.ring,
> +			    &map->active.data, &msg->msg_iter,
> +			    len);
> +	if (sent > 0) {
> +		len -= sent;
> +		tot_sent += sent;
> +		notify_remote_via_irq(map->active.irq);
> +	}
> +	if (sent >= 0 && len > 0 && count < PVCALLS_FRONT_MAX_SPIN)
> +		goto again;
> +	if (sent < 0)
> +		tot_sent = sent;

What does it mean when an error is detected on the interface? Does it 
need to be somehow propagated to the caller?

-boris

> +
> +	mutex_unlock(&map->active.out_mutex);
> +	return tot_sent;
> +}
> +
>   int pvcalls_front_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
>   {
>   	struct pvcalls_bedata *bedata;
> diff --git a/drivers/xen/pvcalls-front.h b/drivers/xen/pvcalls-front.h
> index ab4f1da..d937c24 100644
> --- a/drivers/xen/pvcalls-front.h
> +++ b/drivers/xen/pvcalls-front.h
> @@ -13,5 +13,8 @@ int pvcalls_front_bind(struct socket *sock,
>   int pvcalls_front_accept(struct socket *sock,
>   			 struct socket *newsock,
>   			 int flags);
> +int pvcalls_front_sendmsg(struct socket *sock,
> +			  struct msghdr *msg,
> +			  size_t len);
>   
>   #endif
>