Hi Ophir,
Some small comments, please see bellow
On Mon, Oct 23, 2017 at 02:21:56PM +0000, Ophir Munk wrote:
> mlx4 NIC is using lkey in big endian format. Save lkey in this format
> in order to avoid conversions during Tx fast path
>
> Signed-off-by: Ophir Munk <ophi...@mellanox.com>
> ---
> drivers/net/mlx4/mlx4_rxtx.c | 16 +++++++---------
> drivers/net/mlx4/mlx4_rxtx.h | 14 ++++++++------
> 2 files changed, 15 insertions(+), 15 deletions(-)
>
> diff --git a/drivers/net/mlx4/mlx4_rxtx.c b/drivers/net/mlx4/mlx4_rxtx.c
> index ae37f9b..4a77be8 100644
> --- a/drivers/net/mlx4/mlx4_rxtx.c
> +++ b/drivers/net/mlx4/mlx4_rxtx.c
> @@ -233,9 +233,9 @@ mlx4_txq_complete(struct txq *txq)
> * Index in memory pool (MP) where to add memory region (MR)
> *
> * @return
> - * Added mr->lkey on success, (uint32_t)-1 on failure.
> + * Added mr->lkey (rte_be32_t) on success, (uint32_t)-1 on failure.
This comment is strange, it always returns an rte_be32_t no matter what.
> */
> -uint32_t mlx4_txq_add_mr(struct txq *txq, struct rte_mempool *mp, uint32_t i)
> +rte_be32_t mlx4_txq_add_mr(struct txq *txq, struct rte_mempool *mp, uint32_t
> i)
> {
> struct ibv_mr *mr;
>
> @@ -260,9 +260,9 @@ uint32_t mlx4_txq_add_mr(struct txq *txq, struct
> rte_mempool *mp, uint32_t i)
> /* Store the new entry. */
> txq->mp2mr[i].mp = mp;
> txq->mp2mr[i].mr = mr;
> - txq->mp2mr[i].lkey = mr->lkey;
> + txq->mp2mr[i].lkey = rte_cpu_to_be_32(mr->lkey);
> DEBUG("%p: new MR lkey for MP \"%s\" (%p): 0x%08" PRIu32,
> - (void *)txq, mp->name, (void *)mp, txq->mp2mr[i].lkey);
> + (void *)txq, mp->name, (void *)mp, mr->lkey);
> return txq->mp2mr[i].lkey;
> }
>
> @@ -289,7 +289,6 @@ mlx4_post_send(struct txq *txq, struct rte_mbuf *pkt)
> uint16_t flags16[2];
> } srcrb;
> uint32_t head_idx = sq->head & sq->txbb_cnt_mask;
> - uint32_t lkey;
> uintptr_t addr;
> uint32_t owner_opcode = MLX4_OPCODE_SEND;
> uint32_t byte_count;
> @@ -323,10 +322,10 @@ mlx4_post_send(struct txq *txq, struct rte_mbuf *pkt)
> if (unlikely(dseg >= (struct mlx4_wqe_data_seg *)sq->eob))
> dseg = (struct mlx4_wqe_data_seg *)sq->buf;
> dseg->addr = rte_cpu_to_be_64(addr);
> - /* Memory region key for this memory pool. */
> - lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(buf));
> + /* Memory region key (big endian) for this memory pool. */
> + dseg->lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(buf));
Is not the comment redundant with the return type of the function?
> #ifndef NDEBUG
> - if (unlikely(lkey == (uint32_t)-1)) {
> + if (unlikely(dseg->lkey == rte_cpu_to_be_32((uint32_t)-1))) {
((uint32_t)-1) == ((rte_be32_t)-1) like (uin32t_t)0 == (rte_be32_t)0.
Why do you need this extra conversion?
> /* MR does not exist. */
> DEBUG("%p: unable to get MP <-> MR association",
> (void *)txq);
> @@ -341,7 +340,6 @@ mlx4_post_send(struct txq *txq, struct rte_mbuf *pkt)
> return -EFAULT;
> }
> #endif /* NDEBUG */
> - dseg->lkey = rte_cpu_to_be_32(lkey);
> if (likely(buf->data_len)) {
> byte_count = rte_cpu_to_be_32(buf->data_len);
> } else {
> diff --git a/drivers/net/mlx4/mlx4_rxtx.h b/drivers/net/mlx4/mlx4_rxtx.h
> index 719ef45..b1e8ac4 100644
> --- a/drivers/net/mlx4/mlx4_rxtx.h
> +++ b/drivers/net/mlx4/mlx4_rxtx.h
> @@ -135,7 +135,7 @@ struct txq {
> struct {
> const struct rte_mempool *mp; /**< Cached memory pool. */
> struct ibv_mr *mr; /**< Memory region (for mp). */
> - uint32_t lkey; /**< mr->lkey copy. */
> + rte_be32_t lkey; /**< mr->lkey copy. (big endian) */
The "big endian" in the comment is redundant.
> } mp2mr[MLX4_PMD_TX_MP_CACHE]; /**< MP to MR translation table. */
> struct priv *priv; /**< Back pointer to private data. */
> unsigned int socket; /**< CPU socket ID for allocations. */
> @@ -169,7 +169,7 @@ uint16_t mlx4_tx_burst_removed(void *dpdk_txq, struct
> rte_mbuf **pkts,
> uint16_t pkts_n);
> uint16_t mlx4_rx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts,
> uint16_t pkts_n);
> -uint32_t mlx4_txq_add_mr(struct txq *txq, struct rte_mempool *mp,
> +rte_be32_t mlx4_txq_add_mr(struct txq *txq, struct rte_mempool *mp,
> unsigned int i);
>
> /* mlx4_txq.c */
> @@ -207,9 +207,9 @@ mlx4_txq_mb2mp(struct rte_mbuf *buf)
> * Memory pool for which a memory region lkey must be returned.
> *
> * @return
> - * mr->lkey on success, (uint32_t)-1 on failure.
> + * mr->lkey (rte_be32_t) on success, (uint32_t)-1 on failure.
The comment is also strange, the return type does not change in the mean time.
> */
> -static __rte_always_inline uint32_t
> +static __rte_always_inline rte_be32_t
> mlx4_txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
> {
> unsigned int i;
> @@ -220,8 +220,10 @@ mlx4_txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
> break;
> }
> if (txq->mp2mr[i].mp == mp) {
> - assert(txq->mp2mr[i].lkey != (uint32_t)-1);
> - assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
> + assert(txq->mp2mr[i].lkey !=
> + rte_cpu_to_be_32((uint32_t)-1));
Thanks,
--
Nélio Laranjeiro
6WIND
