On Sun, Jan 01, 2017 at 02:01:57PM -0500, Charles (Chas) Williams wrote:
> .dev_start()/.dev_stop() roughly corresponds to the local device's
> port being up or down. This is different from the remote client being
> connected which is roughtly link up or down. Emulate the behavior by
> separately tracking the local start/stop state to determine if we should
> allow packets to be queued to the remote client.
>
> Signed-off-by: Chas Williams <ciwil...@brocade.com>
> ---
> drivers/net/vhost/rte_eth_vhost.c | 65
> ++++++++++++++++++++++++++++++++-------
> 1 file changed, 54 insertions(+), 11 deletions(-)
>
> diff --git a/drivers/net/vhost/rte_eth_vhost.c
> b/drivers/net/vhost/rte_eth_vhost.c
> index 6b11e40..d5a4540 100644
> --- a/drivers/net/vhost/rte_eth_vhost.c
> +++ b/drivers/net/vhost/rte_eth_vhost.c
> @@ -100,7 +100,8 @@ struct vhost_stats {
>
> struct vhost_queue {
> int vid;
> - rte_atomic32_t allow_queuing;
> + rte_atomic32_t connected;
> + rte_atomic32_t ready;
> rte_atomic32_t while_queuing;
> struct pmd_internal *internal;
> struct rte_mempool *mb_pool;
> @@ -383,18 +384,25 @@ vhost_update_packet_xstats(struct vhost_queue *vq,
> }
> }
>
> +static inline bool
> +queuing_stopped(struct vhost_queue *r)
> +{
> + return unlikely(rte_atomic32_read(&r->connected) == 0 ||
> + rte_atomic32_read(&r->ready) == 0);
> +}
That's one more check comparing to the old code, meaning a bit more
expensive than before.
I think we could maintain the same effort by:
- introduce per-device "started" flag: set/unset on dev_start/stop,
respectively.
- introduce per-device "dev_attached" flag: set/unset on
new/destory_device(), respectively.
On update of each flag, setting "allow_queuing" properly.
Okay to you?
--yliu
