On Thu, Mar 18, 2021 at 12:52 PM Lance Richardson wrote: > > Remove early buffer posting logic from burst receive loop to address > several issues: > - Posting receive descriptors without first posting completion > entries risks overflowing the completion queue. > - Posting receive descriptors without updating rx_raw_prod > creates the possibility that the receive descriptor doorbell > can be written twice with the same value. > - Having this logic in the inner descriptor processing loop > can impact performance. > > Fixes: 637e34befd9c ("net/bnxt: optimize Rx processing") > Fixes: 04067844a3e9 ("net/bnxt: reduce CQ queue size without aggregation ring") > Cc: stable@dpdk.org > > Signed-off-by: Lance Richardson > Reviewed-by: Ajit Kumar Khaparde Patch applied to dpdk-next-net-brcm. > --- > drivers/net/bnxt/bnxt_rxr.c | 3 --- > drivers/net/bnxt/bnxt_rxr.h | 2 -- > 2 files changed, 5 deletions(-) > > diff --git a/drivers/net/bnxt/bnxt_rxr.c b/drivers/net/bnxt/bnxt_rxr.c > index c72545ada7..7179c6cb30 100644 > --- a/drivers/net/bnxt/bnxt_rxr.c > +++ b/drivers/net/bnxt/bnxt_rxr.c > @@ -1018,9 +1018,6 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, > raw_cons = NEXT_RAW_CMP(raw_cons); > if (nb_rx_pkts == nb_pkts || nb_rep_rx_pkts == nb_pkts || evt) > break; > - /* Post some Rx buf early in case of larger burst processing */ > - if (nb_rx_pkts == BNXT_RX_POST_THRESH) > - bnxt_db_write(&rxr->rx_db, rxr->rx_raw_prod); > } > > cpr->cp_raw_cons = raw_cons; > diff --git a/drivers/net/bnxt/bnxt_rxr.h b/drivers/net/bnxt/bnxt_rxr.h > index a6fdd7767a..b43256e03e 100644 > --- a/drivers/net/bnxt/bnxt_rxr.h > +++ b/drivers/net/bnxt/bnxt_rxr.h > @@ -41,8 +41,6 @@ static inline uint16_t bnxt_tpa_start_agg_id(struct bnxt *bp, > (((cmp)->agg_bufs_v1 & RX_PKT_CMPL_AGG_BUFS_MASK) >> \ > RX_PKT_CMPL_AGG_BUFS_SFT) > > -#define BNXT_RX_POST_THRESH 32 > - > /* Number of descriptors to process per inner loop in vector mode. */ > #define RTE_BNXT_DESCS_PER_LOOP 4U > > -- > 2.25.1 >