Re: [PATCH v2 4/7] baseband/acc: allocate FCW memory separately

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

On 9/21/23 22:43, Nicolas Chautru wrote:
> This allows more flexibility to the FCW size for the
> unified driver. No actual functional change.
> 
> Signed-off-by: Nicolas Chautru <nicolas.chautru@intel.com>
> ---
>   drivers/baseband/acc/acc_common.h  |  4 +++-
>   drivers/baseband/acc/rte_vrb_pmd.c | 25 ++++++++++++++++++++++++-
>   2 files changed, 27 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/baseband/acc/acc_common.h b/drivers/baseband/acc/acc_common.h
> index df18506e75..b5ee113faf 100644
> --- a/drivers/baseband/acc/acc_common.h
> +++ b/drivers/baseband/acc/acc_common.h
> @@ -101,6 +101,7 @@
>   #define ACC_NUM_QGRPS_PER_WORD         8
>   #define ACC_MAX_NUM_QGRPS              32
>   #define ACC_RING_SIZE_GRANULARITY      64
> +#define ACC_MAX_FCW_SIZE              128
>   
>   /* Constants from K0 computation from 3GPP 38.212 Table 5.4.2.1-2 */
>   #define ACC_N_ZC_1 66 /* N = 66 Zc for BG 1 */
> @@ -582,13 +583,14 @@ struct __rte_cache_aligned acc_queue {
>   	uint32_t aq_enqueued;  /* Count how many "batches" have been enqueued */
>   	uint32_t aq_dequeued;  /* Count how many "batches" have been dequeued */
>   	uint32_t irq_enable;  /* Enable ops dequeue interrupts if set to 1 */
> -	struct rte_mempool *fcw_mempool;  /* FCW mempool */
>   	enum rte_bbdev_op_type op_type;  /* Type of this Queue: TE or TD */
>   	/* Internal Buffers for loopback input */
>   	uint8_t *lb_in;
>   	uint8_t *lb_out;
> +	uint8_t *fcw_ring;
>   	rte_iova_t lb_in_addr_iova;
>   	rte_iova_t lb_out_addr_iova;
> +	rte_iova_t fcw_ring_addr_iova;
>   	int8_t *derm_buffer; /* interim buffer for de-rm in SDK */
>   	struct acc_device *d;
>   };
> diff --git a/drivers/baseband/acc/rte_vrb_pmd.c b/drivers/baseband/acc/rte_vrb_pmd.c
> index 6898a0f802..f460e9ea2a 100644
> --- a/drivers/baseband/acc/rte_vrb_pmd.c
> +++ b/drivers/baseband/acc/rte_vrb_pmd.c
> @@ -883,6 +883,25 @@ vrb_queue_setup(struct rte_bbdev *dev, uint16_t queue_id,
>   		goto free_companion_ring_addr;
>   	}
>   
> +	q->fcw_ring = rte_zmalloc_socket(dev->device->driver->name,
> +			ACC_MAX_FCW_SIZE * d->sw_ring_max_depth,
> +			RTE_CACHE_LINE_SIZE, conf->socket);
> +	if (q->fcw_ring == NULL) {
> +		rte_bbdev_log(ERR, "Failed to allocate fcw_ring memory");
> +		ret = -ENOMEM;
> +		goto free_companion_ring_addr;
> +	}
> +	q->fcw_ring_addr_iova = rte_malloc_virt2iova(q->fcw_ring);

I see this is not done for other rte_malloc_virt2iova(), but this API 
may returns an error (RTE_BAD_IOVA), so it should be checked.

> +
> +	/* For FFT we need to store the FCW separately */
> +	if (conf->op_type == RTE_BBDEV_OP_FFT) {
> +		for (desc_idx = 0; desc_idx < d->sw_ring_max_depth; desc_idx++) {
> +			desc = q->ring_addr + desc_idx;
> +			desc->req.data_ptrs[0].address = q->fcw_ring_addr_iova +
> +					desc_idx * ACC_MAX_FCW_SIZE;
> +		}
> +	}
> +
>   	q->qgrp_id = (q_idx >> VRB1_GRP_ID_SHIFT) & 0xF;
>   	q->vf_id = (q_idx >> VRB1_VF_ID_SHIFT)  & 0x3F;
>   	q->aq_id = q_idx & 0xF;
> @@ -994,6 +1013,7 @@ vrb_queue_release(struct rte_bbdev *dev, uint16_t q_id)
>   	if (q != NULL) {
>   		/* Mark the Queue as un-assigned. */
>   		d->q_assigned_bit_map[q->qgrp_id] &= (~0ULL - (1 << (uint64_t) q->aq_id));
> +		rte_free(q->fcw_ring);
>   		rte_free(q->companion_ring_addr);
>   		rte_free(q->lb_in);
>   		rte_free(q->lb_out);
> @@ -3224,7 +3244,10 @@ vrb_enqueue_fft_one_op(struct acc_queue *q, struct rte_bbdev_fft_op *op,
>   	output = op->fft.base_output.data;
>   	in_offset = op->fft.base_input.offset;
>   	out_offset = op->fft.base_output.offset;
> -	fcw = &desc->req.fcw_fft;
> +
> +	fcw = (struct acc_fcw_fft *) (q->fcw_ring +
> +			((q->sw_ring_head + total_enqueued_cbs) & q->sw_ring_wrap_mask)
> +			* ACC_MAX_FCW_SIZE);
>   
>   	vrb1_fcw_fft_fill(op, fcw);
>   	vrb1_dma_desc_fft_fill(op, &desc->req, input, output, &in_offset, &out_offset);

With above suggested check added, the patch looks good to me.

Thanks,
Maxime