Re: [PATCH v2 07/37] baseband/acc100: add LDPC encoder padding function

[Maxime Coquelin <maxime.coquelin@redhat.com>]

On 8/20/22 04:31, Hernan Vargas wrote:
> LDPC Encoder input may need to be padded to avoid small beat for ACC100.
> Padding 5GDL input buffer length (BLEN) to avoid case (BLEN % 64) <= 8.
> Adding protection for corner case to avoid for 5GDL occurrence of last
> beat within the ACC100 fabric with <= 8B which might trigger a fabric
> corner case hang issue.

Again, it looks like a bug fix.

> Signed-off-by: Hernan Vargas <hernan.vargas@intel.com>
> ---
>   drivers/baseband/acc100/rte_acc100_pmd.c | 28 ++++++++++++++----------
>   1 file changed, 16 insertions(+), 12 deletions(-)
> 
> diff --git a/drivers/baseband/acc100/rte_acc100_pmd.c b/drivers/baseband/acc100/rte_acc100_pmd.c
> index cc7d146e74..4849d822d1 100644
> --- a/drivers/baseband/acc100/rte_acc100_pmd.c
> +++ b/drivers/baseband/acc100/rte_acc100_pmd.c
> @@ -1304,7 +1304,6 @@ acc100_fcw_td_fill(const struct rte_bbdev_dec_op *op, struct acc100_fcw_td *fcw)
>   			RTE_BBDEV_TURBO_HALF_ITERATION_EVEN);
>   }
>   
> -#ifdef RTE_LIBRTE_BBDEV_DEBUG
>   
>   static inline bool
>   is_acc100(struct acc100_queue *q)
> @@ -1317,7 +1316,6 @@ validate_op_required(struct acc100_queue *q)
>   {
>   	return is_acc100(q);
>   }
> -#endif
>   
>   /* Fill in a frame control word for LDPC decoding. */
>   static inline void
> @@ -1773,12 +1771,24 @@ acc100_dma_desc_te_fill(struct rte_bbdev_enc_op *op,
>   	return 0;
>   }
>   
> +/* May need to pad LDPC Encoder input to avoid small beat for ACC100 */
> +static inline uint16_t
> +pad_le_in(uint16_t blen, struct acc100_queue *q)
> +{
> +	if (!is_acc100(q))
> +		return blen;
> +	uint16_t last_beat = blen % 64;

Do not mix variables declaration & code.

> +	if ((last_beat > 0) && (last_beat <= 8))
> +		blen += 8;
> +	return blen;
> +}
> +
>   static inline int
>   acc100_dma_desc_le_fill(struct rte_bbdev_enc_op *op,
>   		struct acc100_dma_req_desc *desc, struct rte_mbuf **input,
>   		struct rte_mbuf *output, uint32_t *in_offset,
>   		uint32_t *out_offset, uint32_t *out_length,
> -		uint32_t *mbuf_total_left, uint32_t *seg_total_left)
> +		uint32_t *mbuf_total_left, uint32_t *seg_total_left, struct acc100_queue *q)
>   {
>   	int next_triplet = 1; /* FCW already done */
>   	uint16_t K, in_length_in_bits, in_length_in_bytes;
> @@ -1802,8 +1812,7 @@ acc100_dma_desc_le_fill(struct rte_bbdev_enc_op *op,
>   	}
>   
>   	next_triplet = acc100_dma_fill_blk_type_in(desc, input, in_offset,
> -			in_length_in_bytes,
> -			seg_total_left, next_triplet);
> +			pad_le_in(in_length_in_bytes, q), seg_total_left, next_triplet);
>   	if (unlikely(next_triplet < 0)) {
>   		rte_bbdev_log(ERR,
>   				"Mismatch between data to process and mbuf data length in bbdev_op: %p",
> @@ -2531,7 +2540,7 @@ enqueue_ldpc_enc_n_op_cb(struct acc100_queue *q, struct rte_bbdev_enc_op **ops,
>   	acc100_header_init(&desc->req);
>   	desc->req.numCBs = num;
>   
> -	in_length_in_bytes = ops[0]->ldpc_enc.input.data->data_len;
> +	in_length_in_bytes = pad_le_in(ops[0]->ldpc_enc.input.data->data_len, q);
>   	out_length = (enc->cb_params.e + 7) >> 3;
>   	desc->req.m2dlen = 1 + num;
>   	desc->req.d2mlen = num;
> @@ -2600,7 +2609,7 @@ enqueue_ldpc_enc_one_op_cb(struct acc100_queue *q, struct rte_bbdev_enc_op *op,
>   
>   	ret = acc100_dma_desc_le_fill(op, &desc->req, &input, output,
>   			&in_offset, &out_offset, &out_length, &mbuf_total_left,
> -			&seg_total_left);
> +			&seg_total_left, q);
>   
>   	if (unlikely(ret < 0))
>   		return ret;
> @@ -3921,9 +3930,6 @@ dequeue_enc_one_op_cb(struct acc100_queue *q, struct rte_bbdev_enc_op **ref_op,
>   
>   	/* Clearing status, it will be set based on response */
>   	op->status = 0;
> -
> -	op->status |= ((rsp.input_err)
> -			? (1 << RTE_BBDEV_DATA_ERROR) : 0);

Is this really related to the purpose of the patch?

>   	op->status |= ((rsp.dma_err) ? (1 << RTE_BBDEV_DRV_ERROR) : 0);
>   	op->status |= ((rsp.fcw_err) ? (1 << RTE_BBDEV_DRV_ERROR) : 0);
>   
> @@ -3994,8 +4000,6 @@ dequeue_enc_one_op_tb(struct acc100_queue *q, struct rte_bbdev_enc_op **ref_op,
>   		rte_bbdev_log_debug("Resp. desc %p: %x", desc,
>   				rsp.val);
>   
> -		op->status |= ((rsp.input_err)
> -				? (1 << RTE_BBDEV_DATA_ERROR) : 0);

Ditto.

>   		op->status |= ((rsp.dma_err) ? (1 << RTE_BBDEV_DRV_ERROR) : 0);
>   		op->status |= ((rsp.fcw_err) ? (1 << RTE_BBDEV_DRV_ERROR) : 0);
>