[RFC PATCH 20/27] net/intel: write descriptors using non-volatile pointers

[Morten Brørup]

> From: Bruce Richardson [mailto:bruce.richardson at intel.com]
> Sent: Friday, 19 December 2025 18.26
> 
> Use a non-volatile uint64_t pointer to store to the descriptor ring.
> This will allow the compiler to optionally merge the stores as it sees
> best.

I suppose there was a reason for the volatile.
Is removing it really safe?
E.g. this will also allow the compiler to reorder stores; not just the pair of 64-bits, but also stores to multiple descriptors.

One more comment inline below.

> 
> Signed-off-by: Bruce Richardson <bruce.richardson at intel.com>
> ---
>  drivers/net/intel/common/tx_scalar_fns.h | 24 ++++++++++++++++--------
>  1 file changed, 16 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/net/intel/common/tx_scalar_fns.h
> b/drivers/net/intel/common/tx_scalar_fns.h
> index 7b643fcf44..95e9acbe60 100644
> --- a/drivers/net/intel/common/tx_scalar_fns.h
> +++ b/drivers/net/intel/common/tx_scalar_fns.h
> @@ -184,6 +184,15 @@ struct ci_timesstamp_queue_fns {
>  	write_ts_tail_t write_ts_tail;
>  };
> 
> +static inline void
> +write_txd(volatile void *txd, uint64_t qw0, uint64_t qw1)
> +{
> +	uint64_t *txd_qw = RTE_CAST_PTR(void *, txd);

If the descriptors are 16-byte aligned, you could mark them as such, so the compiler can use 128-bit stores on architectures where alignment matters.

> +
> +	txd_qw[0] = rte_cpu_to_le_64(qw0);
> +	txd_qw[1] = rte_cpu_to_le_64(qw1);
> +}
> +
>  static inline uint16_t
>  ci_xmit_pkts(struct ci_tx_queue *txq,
>  	     struct rte_mbuf **tx_pkts,
> @@ -313,8 +322,7 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
>  				txe->mbuf = NULL;
>  			}
> 
> -			ctx_txd[0] = cd_qw0;
> -			ctx_txd[1] = cd_qw1;
> +			write_txd(ctx_txd, cd_qw0, cd_qw1);
> 
>  			txe->last_id = tx_last;
>  			tx_id = txe->next_id;
> @@ -361,12 +369,12 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
> 
>  			while ((ol_flags & (RTE_MBUF_F_TX_TCP_SEG |
> RTE_MBUF_F_TX_UDP_SEG)) &&
>  					unlikely(slen > CI_MAX_DATA_PER_TXD)) {
> -				txd->buffer_addr =
> rte_cpu_to_le_64(buf_dma_addr);
> -				txd->cmd_type_offset_bsz =
> rte_cpu_to_le_64(CI_TX_DESC_DTYPE_DATA |
> +				const uint64_t cmd_type_offset_bsz =
> CI_TX_DESC_DTYPE_DATA |
>  					((uint64_t)td_cmd << CI_TXD_QW1_CMD_S) |
>  					((uint64_t)td_offset <<
> CI_TXD_QW1_OFFSET_S) |
>  					((uint64_t)CI_MAX_DATA_PER_TXD <<
> CI_TXD_QW1_TX_BUF_SZ_S) |
> -					((uint64_t)td_tag <<
> CI_TXD_QW1_L2TAG1_S));
> +					((uint64_t)td_tag <<
> CI_TXD_QW1_L2TAG1_S);
> +				write_txd(txd, buf_dma_addr,
> cmd_type_offset_bsz);
> 
>  				buf_dma_addr += CI_MAX_DATA_PER_TXD;
>  				slen -= CI_MAX_DATA_PER_TXD;
> @@ -382,12 +390,12 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
>  			if (m_seg->next == NULL)
>  				td_cmd |= CI_TX_DESC_CMD_EOP;
> 
> -			txd->buffer_addr = rte_cpu_to_le_64(buf_dma_addr);
> -			txd->cmd_type_offset_bsz =
> rte_cpu_to_le_64(CI_TX_DESC_DTYPE_DATA |
> +			const uint64_t cmd_type_offset_bsz =
> CI_TX_DESC_DTYPE_DATA |
>  				((uint64_t)td_cmd << CI_TXD_QW1_CMD_S) |
>  				((uint64_t)td_offset << CI_TXD_QW1_OFFSET_S) |
>  				((uint64_t)slen << CI_TXD_QW1_TX_BUF_SZ_S) |
> -				((uint64_t)td_tag << CI_TXD_QW1_L2TAG1_S));
> +				((uint64_t)td_tag << CI_TXD_QW1_L2TAG1_S);
> +			write_txd(txd, buf_dma_addr, cmd_type_offset_bsz);
> 
>  			txe->last_id = tx_last;
>  			tx_id = txe->next_id;
> --
> 2.51.0