net/virtio: add vectorized packed ring NEON Tx
authorJoyce Kong <joyce.kong@arm.com>
Tue, 17 Nov 2020 10:06:34 +0000 (18:06 +0800)
committerFerruh Yigit <ferruh.yigit@intel.com>
Fri, 8 Jan 2021 17:07:55 +0000 (18:07 +0100)
Optimize packed ring Tx batch path with NEON instructions.

Signed-off-by: Joyce Kong <joyce.kong@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
drivers/net/virtio/virtio_rxtx_packed.h
drivers/net/virtio/virtio_rxtx_packed_neon.h

index 8f5198a..016b6fb 100644 (file)
@@ -28,6 +28,8 @@
 /* flag bits offset in packed ring desc from ID */
 #define FLAGS_BITS_OFFSET ((offsetof(struct vring_packed_desc, flags) - \
        offsetof(struct vring_packed_desc, id)) * BYTE_SIZE)
+#define FLAGS_LEN_BITS_OFFSET ((offsetof(struct vring_packed_desc, flags) - \
+       offsetof(struct vring_packed_desc, len)) * BYTE_SIZE)
 #endif
 
 #define PACKED_FLAGS_MASK ((0ULL | VRING_PACKED_DESC_F_AVAIL_USED) << \
 /* reference count offset in mbuf rearm data */
 #define REFCNT_BITS_OFFSET ((offsetof(struct rte_mbuf, refcnt) - \
        offsetof(struct rte_mbuf, rearm_data)) * BYTE_SIZE)
+
+#ifdef CC_AVX512_SUPPORT
 /* segment number offset in mbuf rearm data */
 #define SEG_NUM_BITS_OFFSET ((offsetof(struct rte_mbuf, nb_segs) - \
        offsetof(struct rte_mbuf, rearm_data)) * BYTE_SIZE)
-
 /* default rearm data */
 #define DEFAULT_REARM_DATA (1ULL << SEG_NUM_BITS_OFFSET | \
        1ULL << REFCNT_BITS_OFFSET)
+#endif
 
 /* id bits offset in packed ring desc higher 64bits */
 #define ID_BITS_OFFSET ((offsetof(struct vring_packed_desc, id) - \
index 8533e1e..01c77b7 100644 (file)
 #include "virtio_rxtx_packed.h"
 #include "virtqueue.h"
 
+static inline int
+virtqueue_enqueue_batch_packed_vec(struct virtnet_tx *txvq,
+                                  struct rte_mbuf **tx_pkts)
+{
+       struct virtqueue *vq = txvq->vq;
+       uint16_t head_size = vq->hw->vtnet_hdr_size;
+       uint16_t idx = vq->vq_avail_idx;
+       struct virtio_net_hdr *hdr;
+       struct vq_desc_extra *dxp;
+       struct vring_packed_desc *p_desc;
+       uint16_t i;
+
+       if (idx & PACKED_BATCH_MASK)
+               return -1;
+
+       if (unlikely((idx + PACKED_BATCH_SIZE) > vq->vq_nentries))
+               return -1;
+
+       /* Map four refcnt and nb_segs from mbufs to one NEON register. */
+       uint8x16_t ref_seg_msk = {
+               2, 3, 4, 5,
+               10, 11, 12, 13,
+               18, 19, 20, 21,
+               26, 27, 28, 29
+       };
+
+       /* Map four data_off from mbufs to one NEON register. */
+       uint8x8_t data_msk = {
+               0, 1,
+               8, 9,
+               16, 17,
+               24, 25
+       };
+
+       uint16x8_t net_hdr_msk = {
+               0xFFFF, 0xFFFF,
+               0, 0, 0, 0
+       };
+
+       uint16x4_t pkts[PACKED_BATCH_SIZE];
+       uint8x16x2_t mbuf;
+       /* Load four mbufs rearm data. */
+       RTE_BUILD_BUG_ON(REFCNT_BITS_OFFSET >= 64);
+       pkts[0] = vld1_u16((uint16_t *)&tx_pkts[0]->rearm_data);
+       pkts[1] = vld1_u16((uint16_t *)&tx_pkts[1]->rearm_data);
+       pkts[2] = vld1_u16((uint16_t *)&tx_pkts[2]->rearm_data);
+       pkts[3] = vld1_u16((uint16_t *)&tx_pkts[3]->rearm_data);
+
+       mbuf.val[0] = vreinterpretq_u8_u16(vcombine_u16(pkts[0], pkts[1]));
+       mbuf.val[1] = vreinterpretq_u8_u16(vcombine_u16(pkts[2], pkts[3]));
+
+       /* refcnt = 1 and nb_segs = 1 */
+       uint32x4_t def_ref_seg = vdupq_n_u32(0x10001);
+       /* Check refcnt and nb_segs. */
+       uint32x4_t ref_seg = vreinterpretq_u32_u8(vqtbl2q_u8(mbuf, ref_seg_msk));
+       poly128_t cmp1 = vreinterpretq_p128_u32(~vceqq_u32(ref_seg, def_ref_seg));
+       if (unlikely(cmp1))
+               return -1;
+
+       /* Check headroom is enough. */
+       uint16x4_t head_rooms = vdup_n_u16(head_size);
+       RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_off) !=
+                        offsetof(struct rte_mbuf, rearm_data));
+       uint16x4_t data_offset = vreinterpret_u16_u8(vqtbl2_u8(mbuf, data_msk));
+       uint64x1_t cmp2 = vreinterpret_u64_u16(vclt_u16(data_offset, head_rooms));
+       if (unlikely(vget_lane_u64(cmp2, 0)))
+               return -1;
+
+       virtio_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               dxp = &vq->vq_descx[idx + i];
+               dxp->ndescs = 1;
+               dxp->cookie = tx_pkts[i];
+       }
+
+       virtio_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               tx_pkts[i]->data_off -= head_size;
+               tx_pkts[i]->data_len += head_size;
+       }
+
+       uint64x2x2_t desc[PACKED_BATCH_SIZE / 2];
+       uint64x2_t base_addr0 = {
+               VIRTIO_MBUF_ADDR(tx_pkts[0], vq) + tx_pkts[0]->data_off,
+               VIRTIO_MBUF_ADDR(tx_pkts[1], vq) + tx_pkts[1]->data_off
+       };
+       uint64x2_t base_addr1 = {
+               VIRTIO_MBUF_ADDR(tx_pkts[2], vq) + tx_pkts[2]->data_off,
+               VIRTIO_MBUF_ADDR(tx_pkts[3], vq) + tx_pkts[3]->data_off
+       };
+
+       desc[0].val[0] = base_addr0;
+       desc[1].val[0] = base_addr1;
+
+       uint64_t flags = (uint64_t)vq->vq_packed.cached_flags << FLAGS_LEN_BITS_OFFSET;
+       uint64x2_t tx_desc0 = {
+               flags | (uint64_t)idx << ID_BITS_OFFSET | tx_pkts[0]->data_len,
+               flags | (uint64_t)(idx + 1) << ID_BITS_OFFSET | tx_pkts[1]->data_len
+       };
+
+       uint64x2_t tx_desc1 = {
+               flags | (uint64_t)(idx + 2) << ID_BITS_OFFSET | tx_pkts[2]->data_len,
+               flags | (uint64_t)(idx + 3) << ID_BITS_OFFSET | tx_pkts[3]->data_len
+       };
+
+       desc[0].val[1] = tx_desc0;
+       desc[1].val[1] = tx_desc1;
+
+       if (!vq->hw->has_tx_offload) {
+               virtio_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+                       hdr = rte_pktmbuf_mtod_offset(tx_pkts[i],
+                                       struct virtio_net_hdr *, -head_size);
+                       /* Clear net hdr. */
+                       uint16x8_t v_hdr = vld1q_u16((void *)hdr);
+                       vst1q_u16((void *)hdr, vandq_u16(v_hdr, net_hdr_msk));
+               }
+       } else {
+               virtio_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+                       hdr = rte_pktmbuf_mtod_offset(tx_pkts[i],
+                                       struct virtio_net_hdr *, -head_size);
+                       virtqueue_xmit_offload(hdr, tx_pkts[i], true);
+               }
+       }
+
+       /* Enqueue packet buffers. */
+       p_desc = &vq->vq_packed.ring.desc[idx];
+       vst2q_u64((uint64_t *)p_desc, desc[0]);
+       vst2q_u64((uint64_t *)(p_desc + 2), desc[1]);
+
+       virtio_update_batch_stats(&txvq->stats, tx_pkts[0]->pkt_len,
+                       tx_pkts[1]->pkt_len, tx_pkts[2]->pkt_len,
+                       tx_pkts[3]->pkt_len);
+
+       vq->vq_avail_idx += PACKED_BATCH_SIZE;
+       vq->vq_free_cnt -= PACKED_BATCH_SIZE;
+
+       if (vq->vq_avail_idx >= vq->vq_nentries) {
+               vq->vq_avail_idx -= vq->vq_nentries;
+               vq->vq_packed.cached_flags ^=
+                       VRING_PACKED_DESC_F_AVAIL_USED;
+       }
+
+       return 0;
+}
+
 static inline int
 virtqueue_dequeue_batch_packed_vec(struct virtnet_rx *rxvq,
                                   struct rte_mbuf **rx_pkts)