From 2cf3151ee49358dba17e5a0b9d1201f7b7e4d4f3 Mon Sep 17 00:00:00 2001 From: Bruce Richardson Date: Wed, 12 Feb 2014 16:44:00 +0000 Subject: [PATCH] virtio: various improvements Signed-off-by: Bruce Richardson --- lib/librte_pmd_virtio/virtio_ethdev.c | 2 +- lib/librte_pmd_virtio/virtio_rxtx.c | 39 ++++++++------ lib/librte_pmd_virtio/virtqueue.h | 75 +++++++++++++++++++++------ 3 files changed, 83 insertions(+), 33 deletions(-) diff --git a/lib/librte_pmd_virtio/virtio_ethdev.c b/lib/librte_pmd_virtio/virtio_ethdev.c index 0c4ecdc912..1cd921b588 100644 --- a/lib/librte_pmd_virtio/virtio_ethdev.c +++ b/lib/librte_pmd_virtio/virtio_ethdev.c @@ -606,7 +606,7 @@ virtio_dev_link_update(struct rte_eth_dev *dev, __rte_unused int wait_to_complet struct rte_eth_link link, old; uint16_t status; struct virtio_hw *hw = - VIRTIO_DEV_PRIVATE_TO_HW(dev->data->dev_private); + VIRTIO_DEV_PRIVATE_TO_HW(dev->data->dev_private); memset(&link, 0, sizeof(link)); virtio_dev_atomic_read_link_status(dev, &link); old = link; diff --git a/lib/librte_pmd_virtio/virtio_rxtx.c b/lib/librte_pmd_virtio/virtio_rxtx.c index fb5d07fae7..fe94a3feae 100644 --- a/lib/librte_pmd_virtio/virtio_rxtx.c +++ b/lib/librte_pmd_virtio/virtio_rxtx.c @@ -37,6 +37,7 @@ #include #include +#include #include #include #include @@ -81,12 +82,14 @@ virtio_dev_vring_start(struct rte_eth_dev *dev, struct virtqueue *vq, int queue_ PMD_INIT_FUNC_TRACE(); /* - * Reinitialise since virtio port might have been stopped and restarted - */ + * 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, vq->vq_alignment); vq->vq_used_cons_idx = 0; vq->vq_desc_head_idx = 0; + vq->vq_avail_idx = 0; + vq->vq_desc_tail_idx = vq->vq_nentries - 1; vq->vq_free_cnt = vq->vq_nentries; memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries); @@ -125,6 +128,7 @@ virtio_dev_vring_start(struct rte_eth_dev *dev, struct virtqueue *vq, int queue_ } nbufs++; } + vq_update_avail_idx(vq); PMD_INIT_LOG(DEBUG, "Allocated %d bufs\n", nbufs); VIRTIO_WRITE_REG_2(vq->hw, VIRTIO_PCI_QUEUE_SEL, VTNET_SQ_RQ_QUEUE_IDX); VIRTIO_WRITE_REG_4(vq->hw, VIRTIO_PCI_QUEUE_PFN, @@ -229,6 +233,7 @@ virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m) } #define VIRTIO_MBUF_BURST_SZ 64 +#define DESC_PER_CACHELINE (CACHE_LINE_SIZE / sizeof(struct vring_desc)) uint16_t virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts) { @@ -247,8 +252,10 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts) 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); + if (likely(num > DESC_PER_CACHELINE)) + num = num - ((rxvq->vq_used_cons_idx + num) % DESC_PER_CACHELINE); if(num == 0) return 0; - num = virtqueue_dequeue_burst(rxvq, rcv_pkts, len, num); + num = virtqueue_dequeue_burst_rx(rxvq, rcv_pkts, len, num); PMD_RX_LOG(DEBUG, "used:%d dequeue:%d\n", nb_used, num); for (i = 0; i < num ; i ++) { rxm = rcv_pkts[i]; @@ -293,6 +300,8 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts) PMD_RX_LOG(DEBUG, "Notified\n"); } } + vq_update_avail_idx(rxvq); + return (nb_rx); } @@ -301,13 +310,11 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts) { struct virtqueue *txvq = tx_queue; struct rte_mbuf *txm; - uint16_t nb_used, nb_tx, count, num, i; + uint16_t nb_used, nb_tx, num; int error; - uint32_t len[VIRTIO_MBUF_BURST_SZ]; - struct rte_mbuf *snd_pkts[VIRTIO_MBUF_BURST_SZ]; struct virtio_hw *hw; - nb_tx = count = 0; + nb_tx = 0; if (unlikely(nb_pkts < 1)) return (nb_pkts); @@ -319,20 +326,17 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts) hw = txvq->hw; num = (uint16_t)(likely(nb_used < VIRTIO_MBUF_BURST_SZ) ? nb_used : VIRTIO_MBUF_BURST_SZ); - num = virtqueue_dequeue_burst(txvq, snd_pkts, len, num); - for (i = 0; i < num ; i ++) { - rte_pktmbuf_free_seg(snd_pkts[i]); - } - while (count++ < nb_pkts) { + while (nb_tx < nb_pkts) { + if (virtqueue_full(txvq) && num) { + virtqueue_dequeue_pkt_tx(txvq); + num--; + } if(!virtqueue_full(txvq)) { txm = tx_pkts[nb_tx]; - /************************************************/ - /***** Enqueue Packet buffers *****/ - /************************************************/ + /* Enqueue Packet buffers */ error = virtqueue_enqueue_xmit(txvq, txm); if (unlikely(error)) { - // rte_pktmbuf_free_seg(txm); /* the upper application will free this packet */ if (error == ENOSPC) PMD_TX_LOG(ERR, "virtqueue_enqueue Free count = 0\n"); else if (error == EMSGSIZE) @@ -345,10 +349,11 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts) hw->eth_stats.obytes += txm->pkt.data_len; } else { PMD_TX_LOG(ERR, "No free tx descriptors to transmit\n"); - virtqueue_notify(txvq); break; } } + vq_update_avail_idx(txvq); + hw->eth_stats.opackets += nb_tx; if(unlikely(virtqueue_kick_prepare(txvq))) { diff --git a/lib/librte_pmd_virtio/virtqueue.h b/lib/librte_pmd_virtio/virtqueue.h index c1cf146d7f..f3c7776b1d 100644 --- a/lib/librte_pmd_virtio/virtqueue.h +++ b/lib/librte_pmd_virtio/virtqueue.h @@ -50,6 +50,12 @@ #define wmb() rte_wmb() #define rmb() rte_rmb() +#ifdef RTE_PMD_PACKET_PREFETCH +#define rte_packet_prefetch(p) rte_prefetch1(p) +#else +#define rte_packet_prefetch(p) do {} while(0) +#endif + #define VIRTQUEUE_MAX_NAME_SZ 32 #define RTE_MBUF_DATA_DMA_ADDR(mb) \ @@ -121,11 +127,13 @@ struct virtqueue { * VQ_RING_DESC_CHAIN_END. */ uint16_t vq_desc_head_idx; + uint16_t vq_desc_tail_idx; /** * Last consumed descriptor in the used table, * trails vq_ring.used->idx. */ uint16_t vq_used_cons_idx; + uint16_t vq_avail_idx; void *virtio_net_hdr_mem; /**< hdr for each xmit packet */ struct vq_desc_extra { @@ -184,20 +192,25 @@ virtqueue_full(const struct virtqueue *vq) #define VIRTQUEUE_NUSED(vq) ((uint16_t)((vq)->vq_ring.used->idx - (vq)->vq_used_cons_idx)) static inline void -vq_ring_update_avail(struct virtqueue *vq, uint16_t desc_idx) +vq_update_avail_idx(struct virtqueue *vq) +{ + rte_compiler_barrier(); + vq->vq_ring.avail->idx = vq->vq_avail_idx; +} + +static inline void +vq_update_avail_ring(struct virtqueue *vq, uint16_t desc_idx) { uint16_t avail_idx; /* * Place the head of the descriptor chain into the next slot and make - * it usable to the host. The chain is made available now rather than - * deferring to virtqueue_notify() in the hopes that if the host is - * currently running on another CPU, we can keep it processing the new - * descriptor. + * it usable to the host. We wait to inform the host until after the burst + * is complete to avoid cache alignment issues with descriptors. This + * also helps to avoid any contention on the available index. */ - avail_idx = (uint16_t)(vq->vq_ring.avail->idx & (vq->vq_nentries - 1)); + avail_idx = (uint16_t)(vq->vq_avail_idx & (vq->vq_nentries - 1)); vq->vq_ring.avail->ring[avail_idx] = desc_idx; - mb(); - vq->vq_ring.avail->idx++; + vq->vq_avail_idx++; } static inline int virtqueue_kick_prepare(struct virtqueue * vq) @@ -219,14 +232,16 @@ virtqueue_notify(struct virtqueue *vq) static inline void vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx) { - struct vring_desc *dp; + struct vring_desc *dp, *dp_tail; struct vq_desc_extra *dxp; + uint16_t desc_idx_last = desc_idx; dp = &vq->vq_ring.desc[desc_idx]; dxp = &vq->vq_descx[desc_idx]; vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs); if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) { while (dp->flags & VRING_DESC_F_NEXT) { + desc_idx_last = dp->next; dp = &vq->vq_ring.desc[dp->next]; } } @@ -237,8 +252,14 @@ vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx) * newly freed chain. If the virtqueue was completely used, then * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above). */ - dp->next = vq->vq_desc_head_idx; - vq->vq_desc_head_idx = desc_idx; + if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) { + vq->vq_desc_head_idx = desc_idx; + } else { + dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx]; + dp_tail->next = desc_idx; + } + vq->vq_desc_tail_idx = desc_idx_last; + dp->next = VQ_RING_DESC_CHAIN_END; } static inline int @@ -271,8 +292,10 @@ virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie) 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_ring_update_avail(vq, head_idx); + vq_update_avail_ring(vq, head_idx); return (0); } @@ -295,6 +318,8 @@ 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); dxp->cookie = (void *)cookie; dxp->ndescs = needed; @@ -308,31 +333,36 @@ virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie) start_dp[idx].flags = 0; idx = start_dp[idx].next; txvq->vq_desc_head_idx = idx; + if (txvq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END) + txvq->vq_desc_tail_idx = idx; txvq->vq_free_cnt = (uint16_t)(txvq->vq_free_cnt - needed); - vq_ring_update_avail(txvq, head_idx); + vq_update_avail_ring(txvq, head_idx); return (0); } static inline uint16_t -virtqueue_dequeue_burst(struct virtqueue *vq, struct rte_mbuf **rx_pkts, uint32_t *len, uint16_t num) +virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts, uint32_t *len, uint16_t num) { struct vring_used_elem *uep; struct rte_mbuf *cookie; uint16_t used_idx, desc_idx; uint16_t i; + /* Caller does the check */ for (i = 0; i < num ; i ++) { used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1)); uep = &vq->vq_ring.used->ring[used_idx]; desc_idx = (uint16_t) uep->id; + len[i] = uep->len; 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", vq->vq_used_cons_idx); break; } - len[i] = uep->len; + rte_prefetch0(cookie); + rte_packet_prefetch(cookie->pkt.data); rx_pkts[i] = cookie; vq->vq_used_cons_idx++; vq_ring_free_chain(vq, desc_idx); @@ -341,6 +371,21 @@ virtqueue_dequeue_burst(struct virtqueue *vq, struct rte_mbuf **rx_pkts, uint32_ return (i); } +static inline uint16_t +virtqueue_dequeue_pkt_tx(struct virtqueue *vq) +{ + struct vring_used_elem *uep; + uint16_t used_idx, desc_idx; + + used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1)); + uep = &vq->vq_ring.used->ring[used_idx]; + desc_idx = (uint16_t) uep->id; + vq->vq_used_cons_idx++; + vq_ring_free_chain(vq, desc_idx); + + return 0; +} + #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP #define VIRTQUEUE_DUMP(vq) do { \ uint16_t used_idx, nused; \ -- 2.20.1