Re: [dpdk-dev] [PATCH v1] ixgbe: remove vector pmd burst size restriction

["Ananyev, Konstantin" <konstantin.ananyev@intel.com>]

> -----Original Message-----
> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Zoltan Kiss
> Sent: Friday, July 31, 2015 11:04 AM
> To: Liang, Cunming; dev@dpdk.org
> Subject: Re: [dpdk-dev] [PATCH v1] ixgbe: remove vector pmd burst size restriction
> 
> 
> 
> 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@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"

Ah yes, missed that when reviewing, that code would be broken if nb_bufs > 32:

        const uint64_t *split_fl64 = (uint64_t *)split_flags;
        if (rxq->pkt_first_seg == NULL &&
                        split_fl64[0] == 0 && split_fl64[1] == 0 &&
                        split_fl64[2] == 0 && split_fl64[3] == 0)
                return nb_bufs; 

right?

Another thing, that I just thought about:
Right now we invoke ixgbe_rxq_rearm() only at the start of _recv_raw_pkts_vec().
Before it was ok, as _recv_raw_pkts_vec() would never try to read more then 32 RXDs.
But what would happen if nb_pkts > rxq->nb_desc and rxq->rxrearm_nb == 0?
I suppose,  _recv_raw_pkts_vec() can wrpa around RXD ring and 'receive' same packet twice?
So we probably better still limit nb_pkts <= 32 at _recv_raw_pkts_vec().

Konstantin

> 
> 
> > @@ -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);
> >