virtio: mergeable buffers
authorOuyang Changchun <changchun.ouyang@intel.com>
Thu, 14 Aug 2014 08:54:35 +0000 (16:54 +0800)
committerThomas Monjalon <thomas.monjalon@6wind.com>
Mon, 25 Aug 2014 15:14:22 +0000 (17:14 +0200)
This patch supports mergeable buffer feature in DPDK based virtio PMD,
which can receive jumbo frame with larger size, like 3K, 4K or even 9K.

Signed-off-by: Changchun Ouyang <changchun.ouyang@intel.com>
Acked-by: Huawei Xie <huawei.xie@intel.com>
Tested-by: Jingguo Fu <jingguox.fu@intel.com>
lib/librte_pmd_virtio/virtio_ethdev.c
lib/librte_pmd_virtio/virtio_ethdev.h
lib/librte_pmd_virtio/virtio_rxtx.c

index 6293ac6..3344ffb 100644 (file)
@@ -337,7 +337,7 @@ int virtio_dev_queue_setup(struct rte_eth_dev *dev,
                snprintf(vq_name, sizeof(vq_name), "port%d_tvq%d_hdrzone",
                        dev->data->port_id, queue_idx);
                vq->virtio_net_hdr_mz = rte_memzone_reserve_aligned(vq_name,
-                       vq_size * sizeof(struct virtio_net_hdr),
+                       vq_size * hw->vtnet_hdr_size,
                        socket_id, 0, CACHE_LINE_SIZE);
                if (vq->virtio_net_hdr_mz == NULL) {
                        rte_free(vq);
@@ -346,7 +346,7 @@ int virtio_dev_queue_setup(struct rte_eth_dev *dev,
                vq->virtio_net_hdr_mem =
                        vq->virtio_net_hdr_mz->phys_addr;
                memset(vq->virtio_net_hdr_mz->addr, 0,
-                       vq_size * sizeof(struct virtio_net_hdr));
+                       vq_size * hw->vtnet_hdr_size);
        } else if (queue_type == VTNET_CQ) {
                /* Allocate a page for control vq command, data and status */
                snprintf(vq_name, sizeof(vq_name), "port%d_cvq_hdrzone",
@@ -571,9 +571,6 @@ virtio_negotiate_features(struct virtio_hw *hw)
        mask |= VIRTIO_NET_F_GUEST_TSO4 | VIRTIO_NET_F_GUEST_TSO6 | VIRTIO_NET_F_GUEST_ECN;
        mask |= VTNET_LRO_FEATURES;
 
-       /* rx_mbuf should not be in multiple merged segments */
-       mask |= VIRTIO_NET_F_MRG_RXBUF;
-
        /* not negotiating INDIRECT descriptor table support */
        mask |= VIRTIO_RING_F_INDIRECT_DESC;
 
@@ -746,7 +743,6 @@ eth_virtio_dev_init(__rte_unused struct eth_driver *eth_drv,
        }
 
        eth_dev->dev_ops = &virtio_eth_dev_ops;
-       eth_dev->rx_pkt_burst = &virtio_recv_pkts;
        eth_dev->tx_pkt_burst = &virtio_xmit_pkts;
 
        if (rte_eal_process_type() == RTE_PROC_SECONDARY)
@@ -801,10 +797,13 @@ eth_virtio_dev_init(__rte_unused struct eth_driver *eth_drv,
        virtio_negotiate_features(hw);
 
        /* Setting up rx_header size for the device */
-       if (vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF))
+       if (vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
+               eth_dev->rx_pkt_burst = &virtio_recv_mergeable_pkts;
                hw->vtnet_hdr_size = sizeof(struct virtio_net_hdr_mrg_rxbuf);
-       else
+       } else {
+               eth_dev->rx_pkt_burst = &virtio_recv_pkts;
                hw->vtnet_hdr_size = sizeof(struct virtio_net_hdr);
+       }
 
        /* Allocate memory for storing MAC addresses */
        eth_dev->data->mac_addrs = rte_zmalloc("virtio", ETHER_ADDR_LEN, 0);
@@ -1009,7 +1008,7 @@ static void virtio_dev_free_mbufs(struct rte_eth_dev *dev)
 
                while ((buf = (struct rte_mbuf *)virtqueue_detatch_unused(
                                        dev->data->rx_queues[i])) != NULL) {
-                       rte_pktmbuf_free_seg(buf);
+                       rte_pktmbuf_free(buf);
                        mbuf_num++;
                }
 
@@ -1028,7 +1027,8 @@ static void virtio_dev_free_mbufs(struct rte_eth_dev *dev)
                mbuf_num = 0;
                while ((buf = (struct rte_mbuf *)virtqueue_detatch_unused(
                                        dev->data->tx_queues[i])) != NULL) {
-                       rte_pktmbuf_free_seg(buf);
+                       rte_pktmbuf_free(buf);
+
                        mbuf_num++;
                }
 
index 858e644..d2e1eed 100644 (file)
@@ -104,6 +104,9 @@ int  virtio_dev_tx_queue_setup(struct rte_eth_dev *dev, uint16_t tx_queue_id,
 uint16_t virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
                uint16_t nb_pkts);
 
+uint16_t virtio_recv_mergeable_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
+               uint16_t nb_pkts);
+
 uint16_t virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
                uint16_t nb_pkts);
 
index fcd8bd1..0b10108 100644 (file)
@@ -146,6 +146,7 @@ static inline int
 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 {
        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;
@@ -165,9 +166,11 @@ virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
        dxp->ndescs = needed;
 
        start_dp = vq->vq_ring.desc;
-       start_dp[idx].addr  =
-               (uint64_t) (cookie->buf_physaddr + RTE_PKTMBUF_HEADROOM - sizeof(struct virtio_net_hdr));
-       start_dp[idx].len   = cookie->buf_len - RTE_PKTMBUF_HEADROOM + sizeof(struct virtio_net_hdr);
+       start_dp[idx].addr =
+               (uint64_t)(cookie->buf_physaddr + 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;
@@ -184,8 +187,10 @@ virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie)
 {
        struct vq_desc_extra *dxp;
        struct vring_desc *start_dp;
-       uint16_t needed = 2;
+       uint16_t seg_num = cookie->pkt.nb_segs;
+       uint16_t needed = 1 + seg_num;
        uint16_t head_idx, idx;
+       uint16_t head_size = txvq->hw->vtnet_hdr_size;
 
        if (unlikely(txvq->vq_free_cnt == 0))
                return -ENOSPC;
@@ -198,19 +203,25 @@ virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie)
        idx = head_idx;
        dxp = &txvq->vq_descx[idx];
        if (dxp->cookie != NULL)
-               rte_pktmbuf_free_seg(dxp->cookie);
+               rte_pktmbuf_free(dxp->cookie);
        dxp->cookie = (void *)cookie;
        dxp->ndescs = needed;
 
        start_dp = txvq->vq_ring.desc;
-       start_dp[idx].addr  =
-               txvq->virtio_net_hdr_mem + idx * sizeof(struct virtio_net_hdr);
-       start_dp[idx].len   = sizeof(struct virtio_net_hdr);
+       start_dp[idx].addr =
+               txvq->virtio_net_hdr_mem + idx * head_size;
+       start_dp[idx].len = (uint32_t)head_size;
        start_dp[idx].flags = VRING_DESC_F_NEXT;
-       idx = start_dp[idx].next;
-       start_dp[idx].addr  = RTE_MBUF_DATA_DMA_ADDR(cookie);
-       start_dp[idx].len   = cookie->pkt.data_len;
-       start_dp[idx].flags = 0;
+
+       for (; ((seg_num > 0) && (cookie != NULL)); seg_num--) {
+               idx = start_dp[idx].next;
+               start_dp[idx].addr  = RTE_MBUF_DATA_DMA_ADDR(cookie);
+               start_dp[idx].len   = cookie->pkt.data_len;
+               start_dp[idx].flags = VRING_DESC_F_NEXT;
+               cookie = cookie->pkt.next;
+       }
+
+       start_dp[idx].flags &= ~VRING_DESC_F_NEXT;
        idx = start_dp[idx].next;
        txvq->vq_desc_head_idx = idx;
        if (txvq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
@@ -284,7 +295,7 @@ virtio_dev_vring_start(struct virtqueue *vq, int queue_type)
                        error = virtqueue_enqueue_recv_refill(vq, m);
 
                        if (error) {
-                               rte_pktmbuf_free_seg(m);
+                               rte_pktmbuf_free(m);
                                break;
                        }
                        nbufs++;
@@ -423,7 +434,7 @@ virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
        error = virtqueue_enqueue_recv_refill(vq, m);
        if (unlikely(error)) {
                RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
-               rte_pktmbuf_free_seg(m);
+               rte_pktmbuf_free(m);
        }
 }
 
@@ -433,13 +444,13 @@ uint16_t
 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 {
        struct virtqueue *rxvq = rx_queue;
-       struct virtio_hw *hw = rxvq->hw;
        struct rte_mbuf *rxm, *new_mbuf;
        uint16_t nb_used, num, nb_rx = 0;
        uint32_t len[VIRTIO_MBUF_BURST_SZ];
        struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
        int error;
        uint32_t i, nb_enqueued = 0;
+       const uint32_t hdr_size = sizeof(struct virtio_net_hdr);
 
        nb_used = VIRTQUEUE_NUSED(rxvq);
 
@@ -460,8 +471,7 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 
                PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 
-               if (unlikely(len[i]
-                            < (uint32_t)hw->vtnet_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(rxvq, rxm);
@@ -471,17 +481,16 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 
                rxm->pkt.in_port = rxvq->port_id;
                rxm->pkt.data = (char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM;
+
                rxm->pkt.nb_segs = 1;
                rxm->pkt.next = NULL;
-               rxm->pkt.pkt_len  = (uint32_t)(len[i]
-                                              - sizeof(struct virtio_net_hdr));
-               rxm->pkt.data_len = (uint16_t)(len[i]
-                                              - sizeof(struct virtio_net_hdr));
+               rxm->pkt.pkt_len = (uint32_t)(len[i] - hdr_size);
+               rxm->pkt.data_len = (uint16_t)(len[i] - hdr_size);
 
                VIRTIO_DUMP_PACKET(rxm, rxm->pkt.data_len);
 
                rx_pkts[nb_rx++] = rxm;
-               rxvq->bytes += len[i] - sizeof(struct virtio_net_hdr);
+               rxvq->bytes += rx_pkts[nb_rx - 1]->pkt.pkt_len;
        }
 
        rxvq->packets += nb_rx;
@@ -498,11 +507,165 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
                }
                error = virtqueue_enqueue_recv_refill(rxvq, new_mbuf);
                if (unlikely(error)) {
-                       rte_pktmbuf_free_seg(new_mbuf);
+                       rte_pktmbuf_free(new_mbuf);
                        break;
                }
                nb_enqueued++;
        }
+
+       if (likely(nb_enqueued)) {
+               if (unlikely(virtqueue_kick_prepare(rxvq))) {
+                       virtqueue_notify(rxvq);
+                       PMD_RX_LOG(DEBUG, "Notified\n");
+               }
+       }
+
+       vq_update_avail_idx(rxvq);
+
+       return nb_rx;
+}
+
+uint16_t
+virtio_recv_mergeable_pkts(void *rx_queue,
+                       struct rte_mbuf **rx_pkts,
+                       uint16_t nb_pkts)
+{
+       struct virtqueue *rxvq = rx_queue;
+       struct rte_mbuf *rxm, *new_mbuf;
+       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 = 0, nb_enqueued = 0;
+       uint32_t seg_num = 0;
+       uint16_t extra_idx = 0;
+       uint32_t seg_res = 0;
+       const uint32_t hdr_size = sizeof(struct virtio_net_hdr_mrg_rxbuf);
+
+       nb_used = VIRTQUEUE_NUSED(rxvq);
+
+       rmb();
+
+       if (nb_used == 0)
+               return 0;
+
+       PMD_RX_LOG(DEBUG, "used:%d\n", nb_used);
+
+       while (i < nb_used) {
+               struct virtio_net_hdr_mrg_rxbuf *header;
+
+               if (nb_rx == nb_pkts)
+                       break;
+
+               num = virtqueue_dequeue_burst_rx(rxvq, rcv_pkts, len, 1);
+               if (num != 1)
+                       continue;
+
+               i++;
+
+               PMD_RX_LOG(DEBUG, "dequeue:%d\n", num);
+               PMD_RX_LOG(DEBUG, "packet len:%d\n", len[0]);
+
+               rxm = rcv_pkts[0];
+
+               if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
+                       PMD_RX_LOG(ERR, "Packet drop\n");
+                       nb_enqueued++;
+                       virtio_discard_rxbuf(rxvq, rxm);
+                       rxvq->errors++;
+                       continue;
+               }
+
+               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;
+
+               rxm->pkt.data = (char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM;
+               rxm->pkt.nb_segs = seg_num;
+               rxm->pkt.next = NULL;
+               rxm->pkt.pkt_len = (uint32_t)(len[0] - hdr_size);
+               rxm->pkt.data_len = (uint16_t)(len[0] - hdr_size);
+
+               rxm->pkt.in_port = rxvq->port_id;
+               rx_pkts[nb_rx] = rxm;
+               prev = rxm;
+
+               seg_res = seg_num - 1;
+
+               while (seg_res != 0) {
+                       /*
+                        * Get extra segments for current uncompleted packet.
+                        */
+                       uint32_t  rcv_cnt =
+                               RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
+                       if (likely(VIRTQUEUE_NUSED(rxvq) >= rcv_cnt)) {
+                               uint32_t rx_num =
+                                       virtqueue_dequeue_burst_rx(rxvq,
+                                       rcv_pkts, len, rcv_cnt);
+                               i += rx_num;
+                               rcv_cnt = rx_num;
+                       } else {
+                               PMD_RX_LOG(ERR,
+                                       "No enough segments for packet.\n");
+                               nb_enqueued++;
+                               virtio_discard_rxbuf(rxvq, rxm);
+                               rxvq->errors++;
+                               break;
+                       }
+
+                       extra_idx = 0;
+
+                       while (extra_idx < rcv_cnt) {
+                               rxm = rcv_pkts[extra_idx];
+
+                               rxm->pkt.data =
+                                       (char *)rxm->buf_addr +
+                                       RTE_PKTMBUF_HEADROOM - hdr_size;
+                               rxm->pkt.next = NULL;
+                               rxm->pkt.pkt_len = (uint32_t)(len[extra_idx]);
+                               rxm->pkt.data_len = (uint16_t)(len[extra_idx]);
+
+                               if (prev)
+                                       prev->pkt.next = rxm;
+
+                               prev = rxm;
+                               rx_pkts[nb_rx]->pkt.pkt_len += rxm->pkt.pkt_len;
+                               extra_idx++;
+                       };
+                       seg_res -= rcv_cnt;
+               }
+
+               VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
+                       rx_pkts[nb_rx]->pkt.data_len);
+
+               rxvq->bytes += rx_pkts[nb_rx]->pkt.pkt_len;
+               nb_rx++;
+       }
+
+       rxvq->packets += nb_rx;
+
+       /* Allocate new mbuf for the used descriptor */
+       error = ENOSPC;
+       while (likely(!virtqueue_full(rxvq))) {
+               new_mbuf = rte_rxmbuf_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(rxvq, new_mbuf);
+               if (unlikely(error)) {
+                       rte_pktmbuf_free(new_mbuf);
+                       break;
+               }
+               nb_enqueued++;
+       }
+
        if (likely(nb_enqueued)) {
                if (unlikely(virtqueue_kick_prepare(rxvq))) {
                        virtqueue_notify(rxvq);
@@ -536,12 +699,16 @@ 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) {
-               if (virtqueue_full(txvq) && num) {
+               int need = tx_pkts[nb_tx]->pkt.nb_segs - txvq->vq_free_cnt;
+               int deq_cnt = RTE_MIN(need, (int)num);
+
+               num -= (deq_cnt > 0) ? deq_cnt : 0;
+               while (deq_cnt > 0) {
                        virtqueue_dequeue_pkt_tx(txvq);
-                       num--;
+                       deq_cnt--;
                }
 
-               if (!virtqueue_full(txvq)) {
+               if (tx_pkts[nb_tx]->pkt.nb_segs <= txvq->vq_free_cnt) {
                        txm = tx_pkts[nb_tx];
                        /* Enqueue Packet buffers */
                        error = virtqueue_enqueue_xmit(txvq, txm);
@@ -555,7 +722,7 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
                                break;
                        }
                        nb_tx++;
-                       txvq->bytes += txm->pkt.data_len;
+                       txvq->bytes += txm->pkt.pkt_len;
                } else {
                        PMD_TX_LOG(ERR, "No free tx descriptors to transmit");
                        break;