[dpdk-dev] [PATCH] net/mlx5: fix split buffer Rx

[Ferruh Yigit]

On 11/1/2021 9:56 AM, Dmitry Kozlyuk wrote:
> Routine to lookup LKey on Rx was assuming that the mbuf address
> always belongs to a single mempool: the one associated with an RxQ
> or the MPRQ mempool. This assumption is false for split buffers case.
> A wrong LKey was looked up, resulting in completion errors.
> Modify lookup routines to lookup LKey in the mbuf->pool
> for non-MPRQ cases both on Rx datapath and on queue initialization.
> 
> Fixes: fec28ca0e3a9 ("net/mlx5: support mempool registration")
> 
> Signed-off-by: Dmitry Kozlyuk <dkozlyuk at nvidia.com>
> Reviewed-by: Matan Azrad <matan at nvidia.com>
> Reviewed-by: Viacheslav Ovsiienko <viacheslavo at nvidia.com>
> ---
>   drivers/net/mlx5/mlx5_rx.c |  5 ++++-
>   drivers/net/mlx5/mlx5_rx.h | 37 +++++++++++++++++++++++++++++++++++--
>   2 files changed, 39 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/net/mlx5/mlx5_rx.c b/drivers/net/mlx5/mlx5_rx.c
> index 258a645314..7296e81534 100644
> --- a/drivers/net/mlx5/mlx5_rx.c
> +++ b/drivers/net/mlx5/mlx5_rx.c
> @@ -356,6 +356,7 @@ mlx5_rxq_initialize(struct mlx5_rxq_data *rxq)
>   		volatile struct mlx5_wqe_data_seg *scat;
>   		uintptr_t addr;
>   		uint32_t byte_count;
> +		uint32_t lkey;
>   
>   		if (mlx5_rxq_mprq_enabled(rxq)) {
>   			struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[i];
> @@ -366,6 +367,7 @@ mlx5_rxq_initialize(struct mlx5_rxq_data *rxq)
>   							 1 << rxq->strd_num_n);
>   			byte_count = (1 << rxq->strd_sz_n) *
>   					(1 << rxq->strd_num_n);
> +			lkey = mlx5_rx_addr2mr(rxq, addr);
>   		} else {
>   			struct rte_mbuf *buf = (*rxq->elts)[i];
>   
> @@ -373,13 +375,14 @@ mlx5_rxq_initialize(struct mlx5_rxq_data *rxq)
>   					rxq->wqes)[i];
>   			addr = rte_pktmbuf_mtod(buf, uintptr_t);
>   			byte_count = DATA_LEN(buf);
> +			lkey = mlx5_rx_mb2mr(rxq, buf);
>   		}
>   		/* scat->addr must be able to store a pointer. */
>   		MLX5_ASSERT(sizeof(scat->addr) >= sizeof(uintptr_t));
>   		*scat = (struct mlx5_wqe_data_seg){
>   			.addr = rte_cpu_to_be_64(addr),
>   			.byte_count = rte_cpu_to_be_32(byte_count),
> -			.lkey = mlx5_rx_addr2mr(rxq, addr),
> +			.lkey = lkey,
>   		};
>   	}
>   	rxq->consumed_strd = 0;
> diff --git a/drivers/net/mlx5/mlx5_rx.h b/drivers/net/mlx5/mlx5_rx.h
> index 4952fe1455..5903776e7a 100644
> --- a/drivers/net/mlx5/mlx5_rx.h
> +++ b/drivers/net/mlx5/mlx5_rx.h
> @@ -266,7 +266,7 @@ uint16_t mlx5_rx_burst_mprq_vec(void *dpdk_rxq, struct rte_mbuf **pkts,
>   static int mlx5_rxq_mprq_enabled(struct mlx5_rxq_data *rxq);
>   
>   /**
> - * Query LKey from a packet buffer for Rx. No need to flush local caches
> + * Query LKey for an address on Rx. No need to flush local caches
>    * as the Rx mempool database entries are valid for the lifetime of the queue.
>    *
>    * @param rxq
> @@ -301,7 +301,40 @@ mlx5_rx_addr2mr(struct mlx5_rxq_data *rxq, uintptr_t addr)
>   				     mr_ctrl, mp, addr);
>   }
>   
> -#define mlx5_rx_mb2mr(rxq, mb) mlx5_rx_addr2mr(rxq, (uintptr_t)((mb)->buf_addr))
> +/**
> + * Query LKey from a packet buffer for Rx. No need to flush local caches
> + * as the Rx mempool database entries are valid for the lifetime of the queue.
> + *
> + * @param rxq
> + *   Pointer to Rx queue structure.
> + * @param mb
> + *   Buffer to search the address of.
> + *
> + * @return
> + *   Searched LKey on success, UINT32_MAX on no match.
> + *   This function always succeeds on valid input.
> + */
> +static __rte_always_inline uint32_t
> +mlx5_rx_mb2mr(struct mlx5_rxq_data *rxq, struct rte_mbuf *mb)
> +{
> +	struct mlx5_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
> +	uintptr_t addr = (uintptr_t)mb->buf_addr;
> +	struct mlx5_rxq_ctrl *rxq_ctrl;
> +	uint32_t lkey;
> +
> +	/* Linear search on MR cache array. */
> +	lkey = mlx5_mr_lookup_lkey(mr_ctrl->cache, &mr_ctrl->mru,
> +				   MLX5_MR_CACHE_N, addr);
> +	if (likely(lkey != UINT32_MAX))
> +		return lkey;
> +	/*
> +	 * Slower search in the mempool database on miss.
> +	 * During queue creation rxq->sh is not yet set, so we use rxq_ctrl.
> +	 */
> +	rxq_ctrl = container_of(rxq, struct mlx5_rxq_ctrl, rxq);
> +	return mlx5_mr_mempool2mr_bh(&rxq_ctrl->priv->sh->cdev->mr_scache,
> +				     mr_ctrl, mb->pool, addr);

Hi Dmitry,

I am getting build errors with this patch, although I can see CI build is passed,
can you please check if I am missing something, build error is:

../drivers/net/mlx5/mlx5_rx.h: In function ‘mlx5_rx_mb2mr’:
../drivers/net/mlx5/mlx5_rx.h:354:47: error: ‘struct mlx5_rxq_ctrl’ has no member named ‘priv’
   354 |         return mlx5_mr_mempool2mr_bh(&rxq_ctrl->priv->sh->cdev->mr_scache,
       |                                               ^~