net/virtio: fix queue setup consistency
[dpdk.git] / drivers / net / virtio / virtio_rxtx.c
index 6e7ff27..a32e322 100644 (file)
 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
        ETH_TXQ_FLAGS_NOOFFLOADS)
 
+int
+virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
+{
+       struct virtnet_rx *rxvq = rxq;
+       struct virtqueue *vq = rxvq->vq;
+
+       return VIRTQUEUE_NUSED(vq) >= offset;
+}
+
 static void
 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
 {
@@ -124,7 +133,7 @@ virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
                cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 
                if (unlikely(cookie == NULL)) {
-                       PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u\n",
+                       PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
                                vq->vq_used_cons_idx);
                        break;
                }
@@ -258,6 +267,12 @@ tx_offload_enabled(struct virtio_hw *hw)
                vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO6);
 }
 
+/* avoid write operation when necessary, to lessen cache issues */
+#define ASSIGN_UNLESS_EQUAL(var, val) do {     \
+       if ((var) != (val))                     \
+               (var) = (val);                  \
+} while (0)
+
 static inline void
 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
                       uint16_t needed, int use_indirect, int can_push)
@@ -286,8 +301,14 @@ virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
                hdr = (struct virtio_net_hdr *)
                        rte_pktmbuf_prepend(cookie, head_size);
                /* if offload disabled, it is not zeroed below, do it now */
-               if (offload == 0)
-                       memset(hdr, 0, head_size);
+               if (offload == 0) {
+                       ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
+                       ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
+                       ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
+                       ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
+                       ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
+                       ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
+               }
        } else if (use_indirect) {
                /* setup tx ring slot to point to indirect
                 * descriptor list stored in reserved region.
@@ -337,9 +358,9 @@ virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
                        break;
 
                default:
-                       hdr->csum_start = 0;
-                       hdr->csum_offset = 0;
-                       hdr->flags = 0;
+                       ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
+                       ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
+                       ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
                        break;
                }
 
@@ -355,9 +376,9 @@ virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
                                cookie->l3_len +
                                cookie->l4_len;
                } else {
-                       hdr->gso_type = 0;
-                       hdr->gso_size = 0;
-                       hdr->hdr_len = 0;
+                       ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
+                       ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
+                       ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
                }
        }
 
@@ -378,154 +399,16 @@ virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
        vq_update_avail_ring(vq, head_idx);
 }
 
-static void
-virtio_dev_vring_start(struct virtqueue *vq)
-{
-       int size = vq->vq_nentries;
-       struct vring *vr = &vq->vq_ring;
-       uint8_t *ring_mem = vq->vq_ring_virt_mem;
-
-       PMD_INIT_FUNC_TRACE();
-
-       /*
-        * Reinitialise since virtio port might have been stopped and restarted
-        */
-       memset(vq->vq_ring_virt_mem, 0, vq->vq_ring_size);
-       vring_init(vr, size, ring_mem, VIRTIO_PCI_VRING_ALIGN);
-       vq->vq_used_cons_idx = 0;
-       vq->vq_desc_head_idx = 0;
-       vq->vq_avail_idx = 0;
-       vq->vq_desc_tail_idx = (uint16_t)(vq->vq_nentries - 1);
-       vq->vq_free_cnt = vq->vq_nentries;
-       memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries);
-
-       vring_desc_init(vr->desc, size);
-
-       /*
-        * Disable device(host) interrupting guest
-        */
-       virtqueue_disable_intr(vq);
-}
-
 void
 virtio_dev_cq_start(struct rte_eth_dev *dev)
 {
        struct virtio_hw *hw = dev->data->dev_private;
 
        if (hw->cvq && hw->cvq->vq) {
-               virtio_dev_vring_start(hw->cvq->vq);
                VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
        }
 }
 
-void
-virtio_dev_rxtx_start(struct rte_eth_dev *dev)
-{
-       /*
-        * Start receive and transmit vrings
-        * -    Setup vring structure for all queues
-        * -    Initialize descriptor for the rx vring
-        * -    Allocate blank mbufs for the each rx descriptor
-        *
-        */
-       uint16_t i;
-       uint16_t desc_idx;
-       struct virtio_hw *hw = dev->data->dev_private;
-
-       PMD_INIT_FUNC_TRACE();
-
-       /* Start rx vring. */
-       for (i = 0; i < dev->data->nb_rx_queues; i++) {
-               struct virtnet_rx *rxvq = dev->data->rx_queues[i];
-               struct virtqueue *vq = rxvq->vq;
-               int error, nbufs;
-               struct rte_mbuf *m;
-
-               virtio_dev_vring_start(vq);
-               if (rxvq->mpool == NULL) {
-                       rte_exit(EXIT_FAILURE,
-                               "Cannot allocate mbufs for rx virtqueue");
-               }
-
-               /* Allocate blank mbufs for the each rx descriptor */
-               nbufs = 0;
-               error = ENOSPC;
-
-               if (hw->use_simple_rxtx) {
-                       for (desc_idx = 0; desc_idx < vq->vq_nentries;
-                            desc_idx++) {
-                               vq->vq_ring.avail->ring[desc_idx] = desc_idx;
-                               vq->vq_ring.desc[desc_idx].flags =
-                                       VRING_DESC_F_WRITE;
-                       }
-               }
-
-               memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
-               for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
-                    desc_idx++) {
-                       vq->sw_ring[vq->vq_nentries + desc_idx] =
-                               &rxvq->fake_mbuf;
-               }
-
-               while (!virtqueue_full(vq)) {
-                       m = rte_mbuf_raw_alloc(rxvq->mpool);
-                       if (m == NULL)
-                               break;
-
-                       /******************************************
-                       *         Enqueue allocated buffers        *
-                       *******************************************/
-                       if (hw->use_simple_rxtx)
-                               error = virtqueue_enqueue_recv_refill_simple(vq, m);
-                       else
-                               error = virtqueue_enqueue_recv_refill(vq, m);
-
-                       if (error) {
-                               rte_pktmbuf_free(m);
-                               break;
-                       }
-                       nbufs++;
-               }
-
-               vq_update_avail_idx(vq);
-
-               PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
-
-               VIRTQUEUE_DUMP(vq);
-       }
-
-       /* Start tx vring. */
-       for (i = 0; i < dev->data->nb_tx_queues; i++) {
-               struct virtnet_tx *txvq = dev->data->tx_queues[i];
-               struct virtqueue *vq = txvq->vq;
-
-               virtio_dev_vring_start(vq);
-               if (hw->use_simple_rxtx) {
-                       uint16_t mid_idx  = vq->vq_nentries >> 1;
-
-                       for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
-                               vq->vq_ring.avail->ring[desc_idx] =
-                                       desc_idx + mid_idx;
-                               vq->vq_ring.desc[desc_idx + mid_idx].next =
-                                       desc_idx;
-                               vq->vq_ring.desc[desc_idx + mid_idx].addr =
-                                       txvq->virtio_net_hdr_mem +
-                                       offsetof(struct virtio_tx_region, tx_hdr);
-                               vq->vq_ring.desc[desc_idx + mid_idx].len =
-                                       vq->hw->vtnet_hdr_size;
-                               vq->vq_ring.desc[desc_idx + mid_idx].flags =
-                                       VRING_DESC_F_NEXT;
-                               vq->vq_ring.desc[desc_idx].flags = 0;
-                       }
-                       for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
-                            desc_idx++)
-                               vq->vq_ring.avail->ring[desc_idx] = desc_idx;
-               }
-
-               VIRTQUEUE_DUMP(vq);
-       }
-}
-
 int
 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
                        uint16_t queue_idx,
@@ -546,13 +429,75 @@ virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
        vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 
        rxvq = &vq->rxq;
-       rxvq->mpool = mp;
        rxvq->queue_id = queue_idx;
-
+       rxvq->mpool = mp;
+       if (rxvq->mpool == NULL) {
+               rte_exit(EXIT_FAILURE,
+                       "Cannot allocate mbufs for rx virtqueue");
+       }
        dev->data->rx_queues[queue_idx] = rxvq;
 
+       return 0;
+}
+
+int
+virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
+{
+       uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
+       struct virtio_hw *hw = dev->data->dev_private;
+       struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
+       struct virtnet_rx *rxvq = &vq->rxq;
+       struct rte_mbuf *m;
+       uint16_t desc_idx;
+       int error, nbufs;
+
+       PMD_INIT_FUNC_TRACE();
+
+       /* Allocate blank mbufs for the each rx descriptor */
+       nbufs = 0;
+
+       if (hw->use_simple_rxtx) {
+               for (desc_idx = 0; desc_idx < vq->vq_nentries;
+                    desc_idx++) {
+                       vq->vq_ring.avail->ring[desc_idx] = desc_idx;
+                       vq->vq_ring.desc[desc_idx].flags =
+                               VRING_DESC_F_WRITE;
+               }
+       }
+
+       memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
+       for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
+            desc_idx++) {
+               vq->sw_ring[vq->vq_nentries + desc_idx] =
+                       &rxvq->fake_mbuf;
+       }
+
+       while (!virtqueue_full(vq)) {
+               m = rte_mbuf_raw_alloc(rxvq->mpool);
+               if (m == NULL)
+                       break;
+
+               /* Enqueue allocated buffers */
+               if (hw->use_simple_rxtx)
+                       error = virtqueue_enqueue_recv_refill_simple(vq, m);
+               else
+                       error = virtqueue_enqueue_recv_refill(vq, m);
+
+               if (error) {
+                       rte_pktmbuf_free(m);
+                       break;
+               }
+               nbufs++;
+       }
+
+       vq_update_avail_idx(vq);
+
+       PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
+
        virtio_rxq_vec_setup(rxvq);
 
+       VIRTQUEUE_DUMP(vq);
+
        return 0;
 }
 
@@ -632,6 +577,44 @@ virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
        return 0;
 }
 
+int
+virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
+                               uint16_t queue_idx)
+{
+       uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
+       struct virtio_hw *hw = dev->data->dev_private;
+       struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
+       uint16_t mid_idx = vq->vq_nentries >> 1;
+       struct virtnet_tx *txvq = &vq->txq;
+       uint16_t desc_idx;
+
+       PMD_INIT_FUNC_TRACE();
+
+       if (hw->use_simple_rxtx) {
+               for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
+                       vq->vq_ring.avail->ring[desc_idx] =
+                               desc_idx + mid_idx;
+                       vq->vq_ring.desc[desc_idx + mid_idx].next =
+                               desc_idx;
+                       vq->vq_ring.desc[desc_idx + mid_idx].addr =
+                               txvq->virtio_net_hdr_mem +
+                               offsetof(struct virtio_tx_region, tx_hdr);
+                       vq->vq_ring.desc[desc_idx + mid_idx].len =
+                               vq->hw->vtnet_hdr_size;
+                       vq->vq_ring.desc[desc_idx + mid_idx].flags =
+                               VRING_DESC_F_NEXT;
+                       vq->vq_ring.desc[desc_idx].flags = 0;
+               }
+               for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
+                    desc_idx++)
+                       vq->vq_ring.avail->ring[desc_idx] = desc_idx;
+       }
+
+       VIRTQUEUE_DUMP(vq);
+
+       return 0;
+}
+
 static void
 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 {
@@ -766,7 +749,7 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 {
        struct virtnet_rx *rxvq = rx_queue;
        struct virtqueue *vq = rxvq->vq;
-       struct virtio_hw *hw;
+       struct virtio_hw *hw = vq->hw;
        struct rte_mbuf *rxm, *new_mbuf;
        uint16_t nb_used, num, nb_rx;
        uint32_t len[VIRTIO_MBUF_BURST_SZ];
@@ -777,20 +760,23 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
        int offload;
        struct virtio_net_hdr *hdr;
 
+       nb_rx = 0;
+       if (unlikely(hw->started == 0))
+               return nb_rx;
+
        nb_used = VIRTQUEUE_NUSED(vq);
 
        virtio_rmb();
 
-       num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
-       num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
+       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);
 
        num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
        PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 
-       hw = vq->hw;
-       nb_rx = 0;
        nb_enqueued = 0;
        hdr_size = hw->vtnet_hdr_size;
        offload = rx_offload_enabled(hw);
@@ -813,8 +799,6 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
                rxm->ol_flags = 0;
                rxm->vlan_tci = 0;
 
-               rxm->nb_segs = 1;
-               rxm->next = NULL;
                rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
                rxm->data_len = (uint16_t)(len[i] - hdr_size);
 
@@ -834,7 +818,7 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 
                rx_pkts[nb_rx++] = rxm;
 
-               rxvq->stats.bytes += rx_pkts[nb_rx - 1]->pkt_len;
+               rxvq->stats.bytes += rxm->pkt_len;
                virtio_update_packet_stats(&rxvq->stats, rxm);
        }
 
@@ -877,7 +861,7 @@ virtio_recv_mergeable_pkts(void *rx_queue,
 {
        struct virtnet_rx *rxvq = rx_queue;
        struct virtqueue *vq = rxvq->vq;
-       struct virtio_hw *hw;
+       struct virtio_hw *hw = vq->hw;
        struct rte_mbuf *rxm, *new_mbuf;
        uint16_t nb_used, num, nb_rx;
        uint32_t len[VIRTIO_MBUF_BURST_SZ];
@@ -891,14 +875,16 @@ virtio_recv_mergeable_pkts(void *rx_queue,
        uint32_t hdr_size;
        int offload;
 
+       nb_rx = 0;
+       if (unlikely(hw->started == 0))
+               return nb_rx;
+
        nb_used = VIRTQUEUE_NUSED(vq);
 
        virtio_rmb();
 
        PMD_RX_LOG(DEBUG, "used:%d", nb_used);
 
-       hw = vq->hw;
-       nb_rx = 0;
        i = 0;
        nb_enqueued = 0;
        seg_num = 0;
@@ -941,7 +927,6 @@ virtio_recv_mergeable_pkts(void *rx_queue,
 
                rxm->data_off = RTE_PKTMBUF_HEADROOM;
                rxm->nb_segs = seg_num;
-               rxm->next = NULL;
                rxm->ol_flags = 0;
                rxm->vlan_tci = 0;
                rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
@@ -986,7 +971,6 @@ virtio_recv_mergeable_pkts(void *rx_queue,
                                rxm = rcv_pkts[extra_idx];
 
                                rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
-                               rxm->next = NULL;
                                rxm->pkt_len = (uint32_t)(len[extra_idx]);
                                rxm->data_len = (uint16_t)(len[extra_idx]);
 
@@ -1050,9 +1034,12 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
        struct virtqueue *vq = txvq->vq;
        struct virtio_hw *hw = vq->hw;
        uint16_t hdr_size = hw->vtnet_hdr_size;
-       uint16_t nb_used, nb_tx;
+       uint16_t nb_used, nb_tx = 0;
        int error;
 
+       if (unlikely(hw->started == 0))
+               return nb_tx;
+
        if (unlikely(nb_pkts < 1))
                return nb_pkts;
 
@@ -1077,7 +1064,8 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
                }
 
                /* optimize ring usage */
-               if (vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) &&
+               if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
+                     vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
                    rte_mbuf_refcnt_read(txm) == 1 &&
                    RTE_MBUF_DIRECT(txm) &&
                    txm->nb_segs == 1 &&