[dpdk-dev] [PATCH v4] virtio: Fix vring entry number issue

Xie, Huawei huawei.xie at intel.com
Tue Oct 14 18:26:25 CEST 2014



> -----Original Message-----
> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Ouyang Changchun
> Sent: Tuesday, October 14, 2014 9:04 AM
> To: dev at dpdk.org
> Subject: [dpdk-dev] [PATCH v4] virtio: Fix vring entry number issue
> 
>   Fix one issue in virtio TX: it needs one more vring descriptor to hold the virtio
>   header when transmitting packets, it is used later to determine whether to free
>   more entries from used vring.
>   It fixes failing to transmit any packet with 1 segment in the circumstance of
> only
>   1 descriptor in the vring free list.
> 
> Signed-off-by: Changchun Ouyang <changchun.ouyang at intel.com>
> Reviewed-by: Olivier Matz <olivier.matz at 6wind.com>
> ---
> -V4 change:
>   Relocate those changelog and annotation here.
> 
> -V3 change:
>   Resolve compilation issue due to mbuf re-structure.
> 
> -V2 change:
>   Refine description, 'vring descriptor' is more accurate than 'vring entry';
>   Accordingly update test condition (it needs one more descriptor for virtio
> header) to
>   put packets to transmit queue.
> 
> -V1 change:
>   Fix the vring descriptor issue in virtio TX, it needs one more vring descriptor to
> hold the
>   virtio header.
> 
>  lib/librte_pmd_virtio/virtio_rxtx.c | 10 ++++++++--
>  1 file changed, 8 insertions(+), 2 deletions(-)
> 
> diff --git a/lib/librte_pmd_virtio/virtio_rxtx.c
> b/lib/librte_pmd_virtio/virtio_rxtx.c
> index 0b10108..ac44c65 100644
> --- a/lib/librte_pmd_virtio/virtio_rxtx.c
> +++ b/lib/librte_pmd_virtio/virtio_rxtx.c
> @@ -699,7 +699,8 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf
> **tx_pkts, uint16_t nb_pkts)
>  	num = (uint16_t)(likely(nb_used < VIRTIO_MBUF_BURST_SZ) ? nb_used :
> VIRTIO_MBUF_BURST_SZ);
> 
>  	while (nb_tx < nb_pkts) {
> -		int need = tx_pkts[nb_tx]->pkt.nb_segs - txvq->vq_free_cnt;
> +		/* Need one more descriptor for virtio header. */
> +		int need = tx_pkts[nb_tx]->nb_segs - txvq->vq_free_cnt + 1;
>  		int deq_cnt = RTE_MIN(need, (int)num);
> 
>  		num -= (deq_cnt > 0) ? deq_cnt : 0;
> @@ -708,7 +709,12 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf
> **tx_pkts, uint16_t nb_pkts)
>  			deq_cnt--;
>  		}
> 
> -		if (tx_pkts[nb_tx]->pkt.nb_segs <= txvq->vq_free_cnt) {
> +		need = (int)tx_pkts[nb_tx]->nb_segs - txvq->vq_free_cnt + 1;
> +		/*
> +		 * Zero or negative value indicates it has enough free
> +		 * decriptors to transmit.
> +		 */
> +		if (need <= 0) {
Could we put a branch hint to favor the fast path?
>  			txm = tx_pkts[nb_tx];
>  			/* Enqueue Packet buffers */
>  			error = virtqueue_enqueue_xmit(txvq, txm);
> --
> 1.8.4.2



More information about the dev mailing list