On 31/07/15 09:17, Cunming Liang wrote:
> The patch removes the restriction of burst size on a constant 32.
>
> On receive side, the burst size floor now aligns to RTE_IXGBE_DESCS_PER_LOOP
> power of 2.
> According to this rule, the burst size less than 4 still won't receive
> anything.
> (Before this change, the burst size less than 32 can't receive anything.)
>
> On transmit side, the max burst size no longer bind with a constant, however
> it still
> require to check the cross tx_rs_thresh violation.
>
> There's no obvious performance drop found on both recv_pkts_vec
> and recv_scattered_pkts_vec on burst size 32.
>
> Signed-off-by: Cunming Liang <cunming.liang at intel.com>
> ---
> drivers/net/ixgbe/ixgbe_rxtx.c | 3 ++-
> drivers/net/ixgbe/ixgbe_rxtx.h | 4 +---
> drivers/net/ixgbe/ixgbe_rxtx_vec.c | 35 ++++++++++++++++-------------------
> 3 files changed, 19 insertions(+), 23 deletions(-)
>
> diff --git a/drivers/net/ixgbe/ixgbe_rxtx.c b/drivers/net/ixgbe/ixgbe_rxtx.c
> index 3f808b3..dbdb761 100644
> --- a/drivers/net/ixgbe/ixgbe_rxtx.c
> +++ b/drivers/net/ixgbe/ixgbe_rxtx.c
> @@ -4008,7 +4008,8 @@ ixgbe_set_rx_function(struct rte_eth_dev *dev)
> */
> } else if (adapter->rx_vec_allowed) {
> PMD_INIT_LOG(DEBUG, "Vector rx enabled, please make sure RX "
> - "burst size no less than 32.");
> + "burst size no less than 4 (port=%d).",
> + dev->data->port_id);
I think it would be better to use RTE_IXGBE_DESCS_PER_LOOP instead of a
constant 4.
>
> dev->rx_pkt_burst = ixgbe_recv_pkts_vec;
> } else if (adapter->rx_bulk_alloc_allowed) {
> diff --git a/drivers/net/ixgbe/ixgbe_rxtx.h b/drivers/net/ixgbe/ixgbe_rxtx.h
> index 113682a..eb931fe 100644
> --- a/drivers/net/ixgbe/ixgbe_rxtx.h
> +++ b/drivers/net/ixgbe/ixgbe_rxtx.h
> @@ -47,9 +47,7 @@
> (uint64_t) ((mb)->buf_physaddr + RTE_PKTMBUF_HEADROOM)
>
> #ifdef RTE_IXGBE_INC_VECTOR
> -#define RTE_IXGBE_VPMD_RX_BURST 32
> -#define RTE_IXGBE_VPMD_TX_BURST 32
> -#define RTE_IXGBE_RXQ_REARM_THRESH RTE_IXGBE_VPMD_RX_BURST
> +#define RTE_IXGBE_RXQ_REARM_THRESH 32
> #define RTE_IXGBE_TX_MAX_FREE_BUF_SZ 64
> #endif
>
> diff --git a/drivers/net/ixgbe/ixgbe_rxtx_vec.c
> b/drivers/net/ixgbe/ixgbe_rxtx_vec.c
> index 1c16dec..b72f817 100644
> --- a/drivers/net/ixgbe/ixgbe_rxtx_vec.c
> +++ b/drivers/net/ixgbe/ixgbe_rxtx_vec.c
> @@ -199,13 +199,11 @@ desc_to_olflags_v(__m128i descs[4], struct rte_mbuf
> **rx_pkts)
> #endif
>
> /*
> - * vPMD receive routine, now only accept (nb_pkts == RTE_IXGBE_VPMD_RX_BURST)
> - * in one loop
> + * vPMD raw receive routine
I would keep some warning there, like "(if nb_pkts <
RTE_IXGBE_DESCS_PER_LOOP, won't receive anything)"
> *
> * Notice:
> - * - nb_pkts < RTE_IXGBE_VPMD_RX_BURST, just return no packet
> - * - nb_pkts > RTE_IXGBE_VPMD_RX_BURST, only scan RTE_IXGBE_VPMD_RX_BURST
> - * numbers of DD bit
> + * - floor align nb_pkts to a RTE_IXGBE_DESC_PER_LOOP power-of-two
> + * - 'nb_pkts < 4' causes 0 packet receiving
Again, RTE_IXGBE_DESCS_PER_LOOP would be better than 4
> * - don't support ol_flags for rss and csum err
> */
> static inline uint16_t
> @@ -240,8 +238,7 @@ _recv_raw_pkts_vec(struct ixgbe_rx_queue *rxq, struct
> rte_mbuf **rx_pkts,
> __m128i dd_check, eop_check;
> #endif /* RTE_NEXT_ABI */
>
> - if (unlikely(nb_pkts < RTE_IXGBE_VPMD_RX_BURST))
> - return 0;
> + nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_IXGBE_DESCS_PER_LOOP);
>
> /* Just the act of getting into the function from the application is
> * going to cost about 7 cycles */
> @@ -310,7 +307,7 @@ _recv_raw_pkts_vec(struct ixgbe_rx_queue *rxq, struct
> rte_mbuf **rx_pkts,
> * D. fill info. from desc to mbuf
> */
> #endif /* RTE_NEXT_ABI */
> - for (pos = 0, nb_pkts_recd = 0; pos < RTE_IXGBE_VPMD_RX_BURST;
> + for (pos = 0, nb_pkts_recd = 0; pos < nb_pkts;
> pos += RTE_IXGBE_DESCS_PER_LOOP,
> rxdp += RTE_IXGBE_DESCS_PER_LOOP) {
> __m128i descs[RTE_IXGBE_DESCS_PER_LOOP];
> @@ -450,13 +447,11 @@ _recv_raw_pkts_vec(struct ixgbe_rx_queue *rxq, struct
> rte_mbuf **rx_pkts,
> }
>
> /*
> - * vPMD receive routine, now only accept (nb_pkts == RTE_IXGBE_VPMD_RX_BURST)
> - * in one loop
> + * vPMD receive routine
Same note here as above
> *
> * Notice:
> - * - nb_pkts < RTE_IXGBE_VPMD_RX_BURST, just return no packet
> - * - nb_pkts > RTE_IXGBE_VPMD_RX_BURST, only scan RTE_IXGBE_VPMD_RX_BURST
> - * numbers of DD bit
> + * - floor align nb_pkts to a RTE_IXGBE_DESC_PER_LOOP power-of-two
> + * - 'nb_pkts < 4' causes 0 packet receiving
> * - don't support ol_flags for rss and csum err
> */
> uint16_t
> @@ -470,12 +465,11 @@ static inline uint16_t
> reassemble_packets(struct ixgbe_rx_queue *rxq, struct rte_mbuf **rx_bufs,
> uint16_t nb_bufs, uint8_t *split_flags)
> {
> - struct rte_mbuf *pkts[RTE_IXGBE_VPMD_RX_BURST]; /*finished pkts*/
> + struct rte_mbuf *pkts[nb_bufs]; /*finished pkts*/
> struct rte_mbuf *start = rxq->pkt_first_seg;
> struct rte_mbuf *end = rxq->pkt_last_seg;
> unsigned pkt_idx, buf_idx;
>
> -
> for (buf_idx = 0, pkt_idx = 0; buf_idx < nb_bufs; buf_idx++) {
> if (end != NULL) {
> /* processing a split packet */
> @@ -535,14 +529,17 @@ reassemble_packets(struct ixgbe_rx_queue *rxq, struct
> rte_mbuf **rx_bufs,
> *
> * Notice:
> * - don't support ol_flags for rss and csum err
> - * - now only accept (nb_pkts == RTE_IXGBE_VPMD_RX_BURST)
> + * - will floor align nb_pkts to a RTE_IXGBE_DESC_PER_LOOP power-of-two
> + * - 'nb_pkts < 4' causes 0 packet receiving
> */
> uint16_t
> ixgbe_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
> uint16_t nb_pkts)
> {
> struct ixgbe_rx_queue *rxq = rx_queue;
> - uint8_t split_flags[RTE_IXGBE_VPMD_RX_BURST] = {0};
> + uint8_t split_flags[nb_pkts];
> +
> + memset(split_flags, 0, nb_pkts);
>
> /* get some new buffers */
> uint16_t nb_bufs = _recv_raw_pkts_vec(rxq, rx_pkts, nb_pkts,
After this _recv_raw_pkts_vec it checks 32 bytes in split_flags (4x8
bytes), that can overrun or miss some flags.
Btw. Bruce just fixed that part in "ixgbe: fix check for split packets"
> @@ -667,8 +664,8 @@ ixgbe_xmit_pkts_vec(void *tx_queue, struct rte_mbuf
> **tx_pkts,
> uint64_t rs = IXGBE_ADVTXD_DCMD_RS|DCMD_DTYP_FLAGS;
> int i;
>
> - if (unlikely(nb_pkts > RTE_IXGBE_VPMD_TX_BURST))
> - nb_pkts = RTE_IXGBE_VPMD_TX_BURST;
> + /* cross rx_thresh boundary is not allowed */
> + nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
>
> if (txq->nb_tx_free < txq->tx_free_thresh)
> ixgbe_tx_free_bufs(txq);
>
