On Fri, May 23, 2014 at 11:04 AM, Daniele Di Proietto
<ddiproie...@vmware.com> wrote:
> Signed-off-by: Daniele Di Proietto <ddiproie...@vmware.com>
> ---
> lib/netdev-dpdk.c | 139
> +++++++++++++++++++++++++++++++-----------------------
> 1 file changed, 80 insertions(+), 59 deletions(-)
>
> diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c
> index ee811eb..0798a18 100644
> --- a/lib/netdev-dpdk.c
> +++ b/lib/netdev-dpdk.c
> @@ -605,107 +605,128 @@ netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct
> ofpbuf **packets, int *c)
> }
>
> inline static void
> -dpdk_queue_pkt(struct netdev_dpdk *dev, int qid,
> - struct rte_mbuf *pkt)
> +dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
> + struct rte_mbuf **pkts, int c)
> {
> struct dpdk_tx_queue *txq = &dev->tx_q[qid];
> uint64_t diff_tsc;
> uint64_t cur_tsc;
> uint32_t nb_tx;
>
> + int i = 0;
> +
> rte_spinlock_lock(&txq->tx_lock);
> - txq->burst_pkts[txq->count++] = pkt;
> - if (txq->count == MAX_TX_QUEUE_LEN) {
> - goto flush;
> - }
> - cur_tsc = rte_get_timer_cycles();
> - if (txq->count == 1) {
> - txq->tsc = cur_tsc;
> - }
> - diff_tsc = cur_tsc - txq->tsc;
> - if (diff_tsc >= DRAIN_TSC) {
> - goto flush;
> - }
> - rte_spinlock_unlock(&txq->tx_lock);
> - return;
> + while (i < c) {
> + int freeslots = MAX_TX_QUEUE_LEN - txq->count;
> + int tocopy = MIN(freeslots, c-i);
>
> -flush:
> - nb_tx = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts, txq->count);
> - if (nb_tx != txq->count) {
> - /* free buffers if we couldn't transmit packets */
> - rte_mempool_put_bulk(dev->dpdk_mp->mp,
> - (void **) &txq->burst_pkts[nb_tx],
> - (txq->count - nb_tx));
> + memcpy(&txq->burst_pkts[txq->count], &pkts[i], tocopy * sizeof
> (struct rte_mbuf *));
> +
> + txq->count += tocopy;
> + i += tocopy;
> +
> + if (txq->count == MAX_TX_QUEUE_LEN) {
> + goto flush;
> + }
> + cur_tsc = rte_get_timer_cycles();
> + if (txq->count == 1) {
> + txq->tsc = cur_tsc;
> + }
> + diff_tsc = cur_tsc - txq->tsc;
> + if (diff_tsc >= DRAIN_TSC) {
> + goto flush;
> + }
> + continue;
> +
> + flush:
> + nb_tx = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts,
> txq->count);
> + if (nb_tx != txq->count) {
> + /* free buffers if we couldn't transmit packets */
> + rte_mempool_put_bulk(dev->dpdk_mp->mp,
> + (void **) &txq->burst_pkts[nb_tx],
> + (txq->count - nb_tx));
> + }
> + txq->count = 0;
> }
> - txq->count = 0;
> rte_spinlock_unlock(&txq->tx_lock);
> }
>
> /* Tx function. Transmit packets indefinitely */
> static void
> -dpdk_do_tx_copy(struct netdev *netdev, char *buf, int size)
> +dpdk_do_tx_copy(struct netdev *netdev, struct ofpbuf ** ofpbufs, int c)
> {
> struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
> - struct rte_mbuf *pkt;
> + struct rte_mbuf *pkts[c];
> + int i;
>
> - pkt = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
> - if (!pkt) {
> - ovs_mutex_lock(&dev->mutex);
> - dev->stats.tx_dropped++;
> - ovs_mutex_unlock(&dev->mutex);
> - return;
> - }
> + for (i = 0; i < c; i++) {
> + int size = ofpbuf_size(ofpbufs[i]);
> + if (size > dev->max_packet_len) {
> + VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
> + (int)size , dev->max_packet_len);
> +
> + ovs_mutex_lock(&dev->mutex);
> + dev->stats.tx_dropped++;
> + ovs_mutex_unlock(&dev->mutex);
> + }
> +
need to skip this packet.
> + pkts[i] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
> +
> + if (!pkts[i]) {
> + ovs_mutex_lock(&dev->mutex);
> + dev->stats.tx_dropped++;
> + ovs_mutex_unlock(&dev->mutex);
> + return;
> + }
>
> - /* We have to do a copy for now */
> - memcpy(pkt->pkt.data, buf, size);
> + /* We have to do a copy for now */
> + memcpy(pkts[i]->pkt.data, ofpbuf_data(ofpbufs[i]), size);
>
> - rte_pktmbuf_data_len(pkt) = size;
> - rte_pktmbuf_pkt_len(pkt) = size;
> + rte_pktmbuf_data_len(pkts[i]) = size;
> + rte_pktmbuf_pkt_len(pkts[i]) = size;
> + }
>
> - dpdk_queue_pkt(dev, NON_PMD_THREAD_TX_QUEUE, pkt);
> + dpdk_queue_pkts(dev, NON_PMD_THREAD_TX_QUEUE, pkts, c);
> dpdk_queue_flush(dev, NON_PMD_THREAD_TX_QUEUE);
> }
>
> static int
> -netdev_dpdk_send(struct netdev *netdev,
> - struct ofpbuf *ofpbuf, bool may_steal)
> +netdev_dpdk_send_batch(struct netdev *netdev,
> + struct ofpbuf **ofpbufs, int c, bool may_steal)
> {
> struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
> int ret;
> + int i;
>
> - if (ofpbuf_size(ofpbuf) > dev->max_packet_len) {
> - VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
> - (int)ofpbuf_size(ofpbuf) , dev->max_packet_len);
> -
> - ovs_mutex_lock(&dev->mutex);
> - dev->stats.tx_dropped++;
> - ovs_mutex_unlock(&dev->mutex);
> -
> - ret = E2BIG;
> - goto out;
> - }
> -
> - if (!may_steal || ofpbuf->source != OFPBUF_DPDK) {
> - dpdk_do_tx_copy(netdev, (char *) ofpbuf_data(ofpbuf),
> ofpbuf_size(ofpbuf));
> + if (!may_steal || ofpbufs[0]->source != OFPBUF_DPDK) {
> + dpdk_do_tx_copy(netdev, ofpbufs, c);
>
> if (may_steal) {
> - ofpbuf_delete(ofpbuf);
> + for (i = 0; i < c; i++) {
> + ofpbuf_delete(ofpbufs[i]);
> + }
> }
> } else {
> int qid;
>
> qid = rte_lcore_id() % NR_QUEUE;
>
> - dpdk_queue_pkt(dev, qid, (struct rte_mbuf *)ofpbuf);
> -
> + /* TODO: warn if any packet bigger than dev->max_packet_len */
I think we have to check this. Better to add it now, and optimize it later.
> + dpdk_queue_pkts(dev, qid, (struct rte_mbuf **)ofpbufs, c);
> }
> ret = 0;
>
> -out:
> return ret;
> }
>
> static int
> +netdev_dpdk_send(struct netdev *netdev,
> + struct ofpbuf *ofpbuf, bool may_steal)
> +{
> + return netdev_dpdk_send_batch(netdev, &ofpbuf, 1, may_steal);
> +}
> +
> +static int
> netdev_dpdk_set_etheraddr(struct netdev *netdev,
> const uint8_t mac[ETH_ADDR_LEN])
> {
> @@ -1122,7 +1143,7 @@ static struct netdev_class netdev_dpdk_class = {
> NULL, /* get_tunnel_config */
>
> netdev_dpdk_send, /* send */
> - NULL, /* send_batch */
> + netdev_dpdk_send_batch, /* send_batch */
> NULL, /* send_wait */
>
> netdev_dpdk_set_etheraddr,
> --
> 2.0.0.rc0
>
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
