On Mon, Sep 28, 2015 at 01:05:24AM +0800, Zhe Tao wrote:
> The vPMD RX function uses the multi-buffer and SSE instructions to
> accelerate the RX speed, but now the pktype cannot be supported by the vPMD
> RX,
> because it will decrease the performance heavily.
>
> Signed-off-by: Zhe Tao <zhe.tao at intel.com>
> ---
> config/common_bsdapp | 2 +
> config/common_linuxapp | 2 +
> drivers/net/i40e/Makefile | 1 +
> drivers/net/i40e/base/i40e_type.h | 3 +
> drivers/net/i40e/i40e_rxtx.c | 28 ++-
> drivers/net/i40e/i40e_rxtx.h | 20 +-
> drivers/net/i40e/i40e_rxtx_vec.c | 484
> ++++++++++++++++++++++++++++++++++++++
> 7 files changed, 535 insertions(+), 5 deletions(-)
> create mode 100644 drivers/net/i40e/i40e_rxtx_vec.c
>
<snip>
> +
> + /* vPMD receive routine, now only accept (nb_pkts == RTE_I40E_VPMD_RX_BURST)
> + * in one loop
> + *
> + * Notice:
> + * - nb_pkts < RTE_I40E_VPMD_RX_BURST, just return no packet
I don't think this comment matches the implementation below. I think you are
allowed to request bursts as small as RTE_I40E_DESCS_PER_LOOP.
> + * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
> + * numbers of DD bits
> +
> + */
> +static inline uint16_t
> +_recv_raw_pkts_vec(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
> + uint16_t nb_pkts, uint8_t *split_packet)
> +{
> + volatile union i40e_rx_desc *rxdp;
> + struct i40e_rx_entry *sw_ring;
> + uint16_t nb_pkts_recd;
> + int pos;
> + uint64_t var;
> + __m128i shuf_msk;
> +
> + __m128i crc_adjust = _mm_set_epi16(
> + 0, 0, 0, /* ignore non-length fields */
> + -rxq->crc_len, /* sub crc on data_len */
> + 0, /* ignore high-16bits of pkt_len */
> + -rxq->crc_len, /* sub crc on pkt_len */
> + 0, 0 /* ignore pkt_type field */
> + );
> + __m128i dd_check, eop_check;
> +
> + /* nb_pkts shall be less equal than RTE_I40E_MAX_RX_BURST */
> + nb_pkts = RTE_MIN(nb_pkts, RTE_I40E_MAX_RX_BURST);
> +
> + /* nb_pkts has to be floor-aligned to RTE_I40E_DESCS_PER_LOOP */
> + nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_I40E_DESCS_PER_LOOP);
/Bruce
