[dpdk-dev] [PATCH 2/2] net/sfc: free mbufs in bulks on simple EF10 Tx datapath reap

[Ferruh Yigit]

On 9/8/2017 3:15 PM, Andrew Rybchenko wrote:
> From: Ivan Malov <ivan.malov at oktetlabs.ru>
> 
> Signed-off-by: Ivan Malov <ivan.malov at oktetlabs.ru>
> Signed-off-by: Andrew Rybchenko <arybchenko at solarflare.com>
> ---
>  doc/guides/nics/sfc_efx.rst   |  4 +++-
>  drivers/net/sfc/sfc_dp_tx.h   |  2 ++
>  drivers/net/sfc/sfc_ef10_tx.c | 15 ++++++++++++++-
>  drivers/net/sfc/sfc_ethdev.c  |  6 ++++++
>  drivers/net/sfc/sfc_tx.c      | 17 +++++++++++++++++
>  5 files changed, 42 insertions(+), 2 deletions(-)
> 
> diff --git a/doc/guides/nics/sfc_efx.rst b/doc/guides/nics/sfc_efx.rst
> index 973a4a0..028b980 100644
> --- a/doc/guides/nics/sfc_efx.rst
> +++ b/doc/guides/nics/sfc_efx.rst
> @@ -245,12 +245,14 @@ boolean parameters value.
>    features available and required by the datapath implementation.
>    **efx** chooses libefx-based datapath which supports VLAN insertion
>    (full-feature firmware variant only), TSO and multi-segment mbufs.
> +  Mbuf segments may come from different mempools, and mbuf reference
> +  counters are treated responsibly.

This is also the case for ef10 native, right? Does it make sense to
document it in below too?

>    **ef10** chooses EF10 (SFN7xxx, SFN8xxx) native datapath which is
>    more efficient than libefx-based but has no VLAN insertion and TSO
>    support yet.
>    **ef10_simple** chooses EF10 (SFN7xxx, SFN8xxx) native datapath which
>    is even more faster then **ef10** but does not support multi-segment
> -  mbufs.
> +  mbufs, disallows multiple mempools and neglects mbuf reference counters.
>  
>  - ``perf_profile`` [auto|throughput|low-latency] (default **throughput**)
>  

<...>

> --- a/drivers/net/sfc/sfc_ef10_tx.c
> +++ b/drivers/net/sfc/sfc_ef10_tx.c
> @@ -401,14 +401,25 @@ struct sfc_ef10_txq {
>  	pending += sfc_ef10_tx_process_events(txq);
>  
>  	if (pending != completed) {
> +		struct rte_mbuf *bulk[SFC_TX_REAP_BULK_SIZE];
> +		unsigned int nb = 0;
> +
>  		do {
>  			struct sfc_ef10_tx_sw_desc *txd;
>  
>  			txd = &txq->sw_ring[completed & ptr_mask];
>  
> -			rte_pktmbuf_free_seg(txd->mbuf);
> +			if (nb == RTE_DIM(bulk)) {
> +				rte_mempool_put_bulk(bulk[0]->pool,
> +						     (void *)bulk, nb);

same warning here, again false positive I think:
error #3656: variable "bulk" may be used before its value is set

The patch to ignore the warning will take care of this one too.

> +				nb = 0;
> +			}
> +
> +			bulk[nb++] = txd->mbuf;
>  		} while (++completed != pending);
>  
> +		rte_mempool_put_bulk(bulk[0]->pool, (void *)bulk, nb);
> +
>  		txq->completed = completed;
>  	}

<...>