On 02/24/2017 08:48 AM, Zhiyong Yang wrote:
> vhost removes limit of TX burst size(32 pkts) and supports to make
> an best effort to transmit pkts.
>
> Cc: yuanhan....@linux.intel.com
> Cc: maxime.coque...@redhat.com
>
> Signed-off-by: Zhiyong Yang <zhiyong.y...@intel.com>
> ---
> drivers/net/vhost/rte_eth_vhost.c | 24 ++++++++++++++++++++++--
> 1 file changed, 22 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/net/vhost/rte_eth_vhost.c
> b/drivers/net/vhost/rte_eth_vhost.c
> index e98cffd..1e1fa34 100644
> --- a/drivers/net/vhost/rte_eth_vhost.c
> +++ b/drivers/net/vhost/rte_eth_vhost.c
> @@ -52,6 +52,7 @@
> #define ETH_VHOST_QUEUES_ARG "queues"
> #define ETH_VHOST_CLIENT_ARG "client"
> #define ETH_VHOST_DEQUEUE_ZERO_COPY "dequeue-zero-copy"
> +#define VHOST_MAX_PKT_BURST 32
>
> static const char *valid_arguments[] = {
> ETH_VHOST_IFACE_ARG,
> @@ -434,8 +435,27 @@ eth_vhost_tx(void *q, struct rte_mbuf **bufs, uint16_t
> nb_bufs)
> goto out;
>
> /* Enqueue packets to guest RX queue */
> - nb_tx = rte_vhost_enqueue_burst(r->vid,
> - r->virtqueue_id, bufs, nb_bufs);
> + if (likely(nb_bufs <= VHOST_MAX_PKT_BURST))
> + nb_tx = rte_vhost_enqueue_burst(r->vid, r->virtqueue_id,
> + bufs, nb_bufs);
> + else {
> + uint16_t nb_send = nb_bufs;
> +
> + while (nb_send) {
> + uint16_t nb_pkts;
> + uint16_t num = (uint16_t)RTE_MIN(nb_send,
> + VHOST_MAX_PKT_BURST);
> +
> + nb_pkts = rte_vhost_enqueue_burst(r->vid,
> + r->virtqueue_id,
> + &bufs[nb_tx], num);
> +
> + nb_tx += nb_pkts;
> + nb_send -= nb_pkts;
> + if (nb_pkts < num)
> + break;
> + }
In the code above,
- if the VM does not service the queue, this will spin forever
- if the queue is almost full, it will be very slow
In the example of OVS, other ports being serviced by the same core would
stop being serviced or drop a lot of packets.
If you want to remove the 32 pkt limitation, the
rte_vhost_enqueue_burst() api limitation needs to be changed first. It
doesn't make sense to put retry code in eth_vhost_tx() fn, the
application can retry if it wants to.
> + }
>
> r->stats.pkts += nb_tx;
> r->stats.missed_pkts += nb_bufs - nb_tx;
>
