net/virtio: improve batching in mergeable path
authorMaxime Coquelin <maxime.coquelin@redhat.com>
Thu, 20 Dec 2018 17:27:18 +0000 (18:27 +0100)
committerFerruh Yigit <ferruh.yigit@intel.com>
Fri, 21 Dec 2018 15:22:41 +0000 (16:22 +0100)
This patch improves both descriptors dequeue and refill,
by using the same batching strategy as done in in-order path.

Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Jens Freimann <jfreimann@redhat.com>
Reviewed-by: Jens Freimann <jfreimann@redhat.com>
Reviewed-by: Gavin Hu <gavin.hu@arm.com>
drivers/net/virtio/virtio_rxtx.c

index 8f90804..2309b71 100644 (file)
@@ -353,41 +353,44 @@ virtqueue_enqueue_refill_inorder(struct virtqueue *vq,
 }
 
 static inline int
-virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
+virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf **cookie,
+                               uint16_t num)
 {
        struct vq_desc_extra *dxp;
        struct virtio_hw *hw = vq->hw;
-       struct vring_desc *start_dp;
-       uint16_t needed = 1;
-       uint16_t head_idx, idx;
+       struct vring_desc *start_dp = vq->vq_ring.desc;
+       uint16_t idx, i;
 
        if (unlikely(vq->vq_free_cnt == 0))
                return -ENOSPC;
-       if (unlikely(vq->vq_free_cnt < needed))
+       if (unlikely(vq->vq_free_cnt < num))
                return -EMSGSIZE;
 
-       head_idx = vq->vq_desc_head_idx;
-       if (unlikely(head_idx >= vq->vq_nentries))
+       if (unlikely(vq->vq_desc_head_idx >= vq->vq_nentries))
                return -EFAULT;
 
-       idx = head_idx;
-       dxp = &vq->vq_descx[idx];
-       dxp->cookie = (void *)cookie;
-       dxp->ndescs = needed;
+       for (i = 0; i < num; i++) {
+               idx = vq->vq_desc_head_idx;
+               dxp = &vq->vq_descx[idx];
+               dxp->cookie = (void *)cookie[i];
+               dxp->ndescs = 1;
 
-       start_dp = vq->vq_ring.desc;
-       start_dp[idx].addr =
-               VIRTIO_MBUF_ADDR(cookie, vq) +
-               RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
-       start_dp[idx].len =
-               cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
-       start_dp[idx].flags =  VRING_DESC_F_WRITE;
-       idx = start_dp[idx].next;
-       vq->vq_desc_head_idx = idx;
-       if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
-               vq->vq_desc_tail_idx = idx;
-       vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
-       vq_update_avail_ring(vq, head_idx);
+               start_dp[idx].addr =
+                       VIRTIO_MBUF_ADDR(cookie[i], vq) +
+                       RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
+               start_dp[idx].len =
+                       cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM +
+                       hw->vtnet_hdr_size;
+               start_dp[idx].flags = VRING_DESC_F_WRITE;
+               vq->vq_desc_head_idx = start_dp[idx].next;
+               vq_update_avail_ring(vq, idx);
+               if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END) {
+                       vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
+                       break;
+               }
+       }
+
+       vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 
        return 0;
 }
@@ -892,7 +895,8 @@ virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
                                error = virtqueue_enqueue_recv_refill_packed(vq,
                                                &m, 1);
                        else
-                               error = virtqueue_enqueue_recv_refill(vq, m);
+                               error = virtqueue_enqueue_recv_refill(vq,
+                                               &m, 1);
                        if (error) {
                                rte_pktmbuf_free(m);
                                break;
@@ -991,7 +995,7 @@ virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
        if (vtpci_packed_queue(vq->hw))
                error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
        else
-               error = virtqueue_enqueue_recv_refill(vq, m);
+               error = virtqueue_enqueue_recv_refill(vq, &m, 1);
 
        if (unlikely(error)) {
                RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
@@ -1211,7 +1215,7 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
                        dev->data->rx_mbuf_alloc_failed++;
                        break;
                }
-               error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
+               error = virtqueue_enqueue_recv_refill(vq, &new_mbuf, 1);
                if (unlikely(error)) {
                        rte_pktmbuf_free(new_mbuf);
                        break;
@@ -1528,19 +1532,18 @@ virtio_recv_mergeable_pkts(void *rx_queue,
        struct virtnet_rx *rxvq = rx_queue;
        struct virtqueue *vq = rxvq->vq;
        struct virtio_hw *hw = vq->hw;
-       struct rte_mbuf *rxm, *new_mbuf;
-       uint16_t nb_used, num, nb_rx;
+       struct rte_mbuf *rxm;
+       struct rte_mbuf *prev;
+       uint16_t nb_used, num, nb_rx = 0;
        uint32_t len[VIRTIO_MBUF_BURST_SZ];
        struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
-       struct rte_mbuf *prev;
        int error;
-       uint32_t i, nb_enqueued;
-       uint32_t seg_num;
-       uint16_t extra_idx;
-       uint32_t seg_res;
-       uint32_t hdr_size;
+       uint32_t nb_enqueued = 0;
+       uint32_t seg_num = 0;
+       uint32_t seg_res = 0;
+       uint32_t hdr_size = hw->vtnet_hdr_size;
+       int32_t i;
 
-       nb_rx = 0;
        if (unlikely(hw->started == 0))
                return nb_rx;
 
@@ -1550,31 +1553,25 @@ virtio_recv_mergeable_pkts(void *rx_queue,
 
        PMD_RX_LOG(DEBUG, "used:%d", nb_used);
 
-       i = 0;
-       nb_enqueued = 0;
-       seg_num = 0;
-       extra_idx = 0;
-       seg_res = 0;
-       hdr_size = hw->vtnet_hdr_size;
-
-       while (i < nb_used) {
-               struct virtio_net_hdr_mrg_rxbuf *header;
+       num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
+       if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
+               num = VIRTIO_MBUF_BURST_SZ;
+       if (likely(num > DESC_PER_CACHELINE))
+               num = num - ((vq->vq_used_cons_idx + num) %
+                               DESC_PER_CACHELINE);
 
-               if (nb_rx == nb_pkts)
-                       break;
 
-               num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
-               if (num != 1)
-                       continue;
+       num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
 
-               i++;
+       for (i = 0; i < num; i++) {
+               struct virtio_net_hdr_mrg_rxbuf *header;
 
                PMD_RX_LOG(DEBUG, "dequeue:%d", num);
-               PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
+               PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 
-               rxm = rcv_pkts[0];
+               rxm = rcv_pkts[i];
 
-               if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
+               if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
                        PMD_RX_LOG(ERR, "Packet drop");
                        nb_enqueued++;
                        virtio_discard_rxbuf(vq, rxm);
@@ -1582,10 +1579,10 @@ virtio_recv_mergeable_pkts(void *rx_queue,
                        continue;
                }
 
-               header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
-                       RTE_PKTMBUF_HEADROOM - hdr_size);
+               header = (struct virtio_net_hdr_mrg_rxbuf *)
+                        ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
+                        - hdr_size);
                seg_num = header->num_buffers;
-
                if (seg_num == 0)
                        seg_num = 1;
 
@@ -1593,10 +1590,11 @@ virtio_recv_mergeable_pkts(void *rx_queue,
                rxm->nb_segs = seg_num;
                rxm->ol_flags = 0;
                rxm->vlan_tci = 0;
-               rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
-               rxm->data_len = (uint16_t)(len[0] - hdr_size);
+               rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
+               rxm->data_len = (uint16_t)(len[i] - hdr_size);
 
                rxm->port = rxvq->port_id;
+
                rx_pkts[nb_rx] = rxm;
                prev = rxm;
 
@@ -1607,75 +1605,96 @@ virtio_recv_mergeable_pkts(void *rx_queue,
                        continue;
                }
 
+               if (hw->vlan_strip)
+                       rte_vlan_strip(rx_pkts[nb_rx]);
+
                seg_res = seg_num - 1;
 
-               while (seg_res != 0) {
-                       /*
-                        * Get extra segments for current uncompleted packet.
-                        */
-                       uint16_t  rcv_cnt =
-                               RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
-                       if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
-                               uint32_t rx_num =
-                                       virtqueue_dequeue_burst_rx(vq,
-                                       rcv_pkts, len, rcv_cnt);
-                               i += rx_num;
-                               rcv_cnt = rx_num;
-                       } else {
-                               PMD_RX_LOG(ERR,
-                                          "No enough segments for packet.");
-                               nb_enqueued++;
-                               virtio_discard_rxbuf(vq, rxm);
-                               rxvq->stats.errors++;
-                               break;
-                       }
+               /* Merge remaining segments */
+               while (seg_res != 0 && i < (num - 1)) {
+                       i++;
+
+                       rxm = rcv_pkts[i];
+                       rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
+                       rxm->pkt_len = (uint32_t)(len[i]);
+                       rxm->data_len = (uint16_t)(len[i]);
+
+                       rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
+                       rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
+
+                       if (prev)
+                               prev->next = rxm;
+
+                       prev = rxm;
+                       seg_res -= 1;
+               }
+
+               if (!seg_res) {
+                       virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
+                       nb_rx++;
+               }
+       }
+
+       /* Last packet still need merge segments */
+       while (seg_res != 0) {
+               uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
+                                       VIRTIO_MBUF_BURST_SZ);
 
-                       extra_idx = 0;
+               prev = rcv_pkts[nb_rx];
+               if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
+                       num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
+                                                          rcv_cnt);
+                       uint16_t extra_idx = 0;
 
+                       rcv_cnt = num;
                        while (extra_idx < rcv_cnt) {
                                rxm = rcv_pkts[extra_idx];
-
-                               rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
+                               rxm->data_off =
+                                       RTE_PKTMBUF_HEADROOM - hdr_size;
                                rxm->pkt_len = (uint32_t)(len[extra_idx]);
                                rxm->data_len = (uint16_t)(len[extra_idx]);
-
-                               if (prev)
-                                       prev->next = rxm;
-
+                               prev->next = rxm;
                                prev = rxm;
-                               rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
-                               extra_idx++;
+                               rx_pkts[nb_rx]->pkt_len += len[extra_idx];
+                               rx_pkts[nb_rx]->data_len += len[extra_idx];
+                               extra_idx += 1;
                        };
                        seg_res -= rcv_cnt;
-               }
-
-               if (hw->vlan_strip)
-                       rte_vlan_strip(rx_pkts[nb_rx]);
-
-               VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
-                       rx_pkts[nb_rx]->data_len);
 
-               virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
-               nb_rx++;
+                       if (!seg_res) {
+                               virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
+                               nb_rx++;
+                       }
+               } else {
+                       PMD_RX_LOG(ERR,
+                                       "No enough segments for packet.");
+                       virtio_discard_rxbuf(vq, prev);
+                       rxvq->stats.errors++;
+                       break;
+               }
        }
 
        rxvq->stats.packets += nb_rx;
 
        /* Allocate new mbuf for the used descriptor */
-       while (likely(!virtqueue_full(vq))) {
-               new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
-               if (unlikely(new_mbuf == NULL)) {
-                       struct rte_eth_dev *dev
-                               = &rte_eth_devices[rxvq->port_id];
-                       dev->data->rx_mbuf_alloc_failed++;
-                       break;
-               }
-               error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
-               if (unlikely(error)) {
-                       rte_pktmbuf_free(new_mbuf);
-                       break;
+       if (likely(!virtqueue_full(vq))) {
+               /* free_cnt may include mrg descs */
+               uint16_t free_cnt = vq->vq_free_cnt;
+               struct rte_mbuf *new_pkts[free_cnt];
+
+               if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
+                       error = virtqueue_enqueue_recv_refill(vq, new_pkts,
+                                       free_cnt);
+                       if (unlikely(error)) {
+                               for (i = 0; i < free_cnt; i++)
+                                       rte_pktmbuf_free(new_pkts[i]);
+                       }
+                       nb_enqueued += free_cnt;
+               } else {
+                       struct rte_eth_dev *dev =
+                               &rte_eth_devices[rxvq->port_id];
+                       dev->data->rx_mbuf_alloc_failed += free_cnt;
                }
-               nb_enqueued++;
        }
 
        if (likely(nb_enqueued)) {