Re: [dpdk-dev] [PATCH v3 5/8] baseband/la12xx: add enqueue and dequeue support

[Hemant Agrawal <hemant.agrawal@oss.nxp.com>]

On 4/14/2021 6:23 AM, Chautru, Nicolas wrote:
> Add support for enqueue and dequeue the LDPC enc/dec from the modem device.
>
> Signed-off-by: Nipun Gupta <nipun.gupta@nxp.com>
> Signed-off-by: Hemant Agrawal <hemant.agrawal@nxp.com>
> ---
>   drivers/baseband/la12xx/bbdev_la12xx.c     | 397 ++++++++++++++++++++-
>   drivers/baseband/la12xx/bbdev_la12xx_ipc.h |  37 ++
>   2 files changed, 430 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/baseband/la12xx/bbdev_la12xx.c b/drivers/baseband/la12xx/bbdev_la12xx.c
> index 0a68686205..d1040987b2 100644
> --- a/drivers/baseband/la12xx/bbdev_la12xx.c
> +++ b/drivers/baseband/la12xx/bbdev_la12xx.c
> @@ -117,6 +117,10 @@ la12xx_queue_release(struct rte_bbdev *dev, uint16_t q_id)
>   		((uint64_t) ((unsigned long) (A) \
>   		- ((uint64_t)ipc_priv->hugepg_start.host_vaddr)))
>   
> +#define MODEM_P2V(A) \
> +	((uint64_t) ((unsigned long) (A) \
> +		+ (unsigned long)(ipc_priv->peb_start.host_vaddr)))
> +
>   static int ipc_queue_configure(uint32_t channel_id,
>   		ipc_t instance, struct bbdev_la12xx_q_priv *q_priv)  { @@ -345,6 +349,387 @@ static const struct rte_bbdev_ops pmd_ops = {
>   	.queue_release = la12xx_queue_release,
>   	.start = la12xx_start
>   };
> +
> +static int
> +fill_feca_desc_enc(struct bbdev_la12xx_q_priv *q_priv,
> +		   struct bbdev_ipc_dequeue_op *bbdev_ipc_op,
> +		   struct rte_bbdev_enc_op *bbdev_enc_op,
> +		   struct rte_bbdev_op_data *in_op_data) {
> +	RTE_SET_USED(q_priv);
> +	RTE_SET_USED(bbdev_ipc_op);
> +	RTE_SET_USED(bbdev_enc_op);
> +	RTE_SET_USED(in_op_data);
> +
> +	return 0;
> +}
>
> I miss why these functions are here.
> Is that contribution supposed to work or a placeholder?

it is currently a placeholder for some optimization work we are doing.

i.e. to covert the bbdev params to our hardware format at host side itself.

We will remove these for now.

>
> +
> +static int
> +fill_feca_desc_dec(struct bbdev_la12xx_q_priv *q_priv,
> +		   struct bbdev_ipc_dequeue_op *bbdev_ipc_op,
> +		   struct rte_bbdev_dec_op *bbdev_dec_op,
> +		   struct rte_bbdev_op_data *out_op_data) {
> +	RTE_SET_USED(q_priv);
> +	RTE_SET_USED(bbdev_ipc_op);
> +	RTE_SET_USED(bbdev_dec_op);
> +	RTE_SET_USED(out_op_data);
> +
> +	return 0;
> +}
> +
> +static inline int
> +is_bd_ring_full(uint32_t ci, uint32_t ci_flag,
> +		uint32_t pi, uint32_t pi_flag)
> +{
> +	if (pi == ci) {
> +		if (pi_flag != ci_flag)
> +			return 1; /* Ring is Full */
> +	}
> +	return 0;
> +}
> +
> +static inline int
> +prepare_ldpc_enc_op(struct rte_bbdev_enc_op *bbdev_enc_op,
> +		    struct bbdev_ipc_dequeue_op *bbdev_ipc_op,
> +		    struct bbdev_la12xx_q_priv *q_priv,
> +		    struct rte_bbdev_op_data *in_op_data,
> +		    struct rte_bbdev_op_data *out_op_data) {
> +	struct rte_bbdev_op_ldpc_enc *ldpc_enc = &bbdev_enc_op->ldpc_enc;
> +	uint32_t total_out_bits;
> +	int ret;
> +
> +	total_out_bits = (ldpc_enc->tb_params.cab *
> +		ldpc_enc->tb_params.ea) + (ldpc_enc->tb_params.c -
> +		ldpc_enc->tb_params.cab) * ldpc_enc->tb_params.eb;
> +
>
> This includes ratematching, see previous comment on capability
>
> Also I see it would not support the partial TB as defined in documentation and API (r != 0)
Not yet.
>
> +	ldpc_enc->output.length = (total_out_bits + 7)/8;
> +
> +	ret = fill_feca_desc_enc(q_priv, bbdev_ipc_op,
> +				 bbdev_enc_op, in_op_data);
> +	if (ret) {
> +		BBDEV_LA12XX_PMD_ERR(
> +			"fill_feca_desc_enc failed, ret: %d", ret);
> +		return ret;
> +	}
> +
> +	rte_pktmbuf_append(out_op_data->data, ldpc_enc->output.length);
> +
> +	return 0;
> +}
> +
> +static inline int
> +prepare_ldpc_dec_op(struct rte_bbdev_dec_op *bbdev_dec_op,
> +		    struct bbdev_ipc_dequeue_op *bbdev_ipc_op,
> +		    struct bbdev_la12xx_q_priv *q_priv,
> +		    struct rte_bbdev_op_data *out_op_data) {
> +	struct rte_bbdev_op_ldpc_dec *ldpc_dec = &bbdev_dec_op->ldpc_dec;
> +	uint32_t total_out_bits;
> +	uint32_t num_code_blocks = 0;
> +	uint16_t sys_cols;
> +	int ret;
> +
> +	sys_cols = (ldpc_dec->basegraph == 1) ? 22 : 10;
> +	if (ldpc_dec->tb_params.c == 1) {
> +		total_out_bits = ((sys_cols * ldpc_dec->z_c) -
> +				ldpc_dec->n_filler);
> +		/* 5G-NR protocol uses 16 bit CRC when output packet
> +		 * size <= 3824 (bits). Otherwise 24 bit CRC is used.
> +		 * Adjust the output bits accordingly
> +		 */
> +		if (total_out_bits - 16 <= 3824)
> +			total_out_bits -= 16;
> +		else
> +			total_out_bits -= 24;
> +		ldpc_dec->hard_output.length = (total_out_bits / 8);
> +	} else {
> +		total_out_bits = (((sys_cols * ldpc_dec->z_c) -
> +				ldpc_dec->n_filler - 24) *
> +				ldpc_dec->tb_params.c);
> +		ldpc_dec->hard_output.length = (total_out_bits / 8) - 3;
>
> Probably good to remove magic number for 24 and 3 here.
ok
>
> +	}
> +
> +	num_code_blocks = ldpc_dec->tb_params.c;
> +
> +	bbdev_ipc_op->num_code_blocks = rte_cpu_to_be_32(num_code_blocks);
> +
> +	ret = fill_feca_desc_dec(q_priv, bbdev_ipc_op,
> +				 bbdev_dec_op, out_op_data);
> +	if (ret) {
> +		BBDEV_LA12XX_PMD_ERR("fill_feca_desc_dec failed, ret: %d", ret);
> +		return ret;
> +	}
> +
> +	return 0;
> +}
> +
> +static int
> +enqueue_single_op(struct bbdev_la12xx_q_priv *q_priv, void *bbdev_op) {
> +	struct bbdev_la12xx_private *priv = q_priv->bbdev_priv;
> +	ipc_userspace_t *ipc_priv = priv->ipc_priv;
> +	ipc_instance_t *ipc_instance = ipc_priv->instance;
> +	struct bbdev_ipc_dequeue_op *bbdev_ipc_op;
> +	struct rte_bbdev_op_ldpc_enc *ldpc_enc;
> +	struct rte_bbdev_op_ldpc_dec *ldpc_dec;
> +	uint32_t q_id = q_priv->q_id;
> +	uint32_t ci, ci_flag, pi, pi_flag;
> +	ipc_ch_t *ch = &(ipc_instance->ch_list[q_id]);
> +	ipc_br_md_t *md = &(ch->md);
> +	size_t virt;
> +	char *huge_start_addr =
> +		(char *)q_priv->bbdev_priv->ipc_priv->hugepg_start.host_vaddr;
> +	struct rte_bbdev_op_data *in_op_data, *out_op_data;
> +	char *data_ptr;
> +	uint32_t l1_pcie_addr;
> +	int ret;
> +	uint32_t temp_ci;
> +
> +	temp_ci = q_priv->host_params->ci;
> +	ci = IPC_GET_CI_INDEX(temp_ci);
> +	ci_flag = IPC_GET_CI_FLAG(temp_ci);
> +
> +	pi = IPC_GET_PI_INDEX(q_priv->host_pi);
> +	pi_flag = IPC_GET_PI_FLAG(q_priv->host_pi);
> +
> +	BBDEV_LA12XX_PMD_DP_DEBUG(
> +		"before bd_ring_full: pi: %u, ci: %u, pi_flag: %u, ci_flag: %u, ring size: %u",
> +		pi, ci, pi_flag, ci_flag, q_priv->queue_size);
> +
> +	if (is_bd_ring_full(ci, ci_flag, pi, pi_flag)) {
> +		BBDEV_LA12XX_PMD_DP_DEBUG(
> +				"bd ring full for queue id: %d", q_id);
> +		return IPC_CH_FULL;
> +	}
> +
> +	virt = MODEM_P2V(q_priv->host_params->modem_ptr[pi]);
> +	bbdev_ipc_op = (struct bbdev_ipc_dequeue_op *)virt;
> +	q_priv->bbdev_op[pi] = bbdev_op;
> +
> +	switch (q_priv->op_type) {
> +	case RTE_BBDEV_OP_LDPC_ENC:
> +		ldpc_enc = &(((struct rte_bbdev_enc_op *)bbdev_op)->ldpc_enc);
> +		in_op_data = &ldpc_enc->input;
> +		out_op_data = &ldpc_enc->output;
> +
> +		ret = prepare_ldpc_enc_op(bbdev_op, bbdev_ipc_op, q_priv,
> +					  in_op_data, out_op_data);
> +		if (ret) {
> +			BBDEV_LA12XX_PMD_ERR(
> +				"process_ldpc_enc_op failed, ret: %d", ret);
> +			return ret;
> +		}
> +		break;
> +
> +	case RTE_BBDEV_OP_LDPC_DEC:
> +		ldpc_dec = &(((struct rte_bbdev_dec_op *)bbdev_op)->ldpc_dec);
> +		in_op_data = &ldpc_dec->input;
> +
> +			out_op_data = &ldpc_dec->hard_output;
> +
> +		ret = prepare_ldpc_dec_op(bbdev_op, bbdev_ipc_op,
> +					  q_priv, out_op_data);
> +		if (ret) {
> +			BBDEV_LA12XX_PMD_ERR(
> +				"process_ldpc_dec_op failed, ret: %d", ret);
> +			return ret;
> +		}
> +		break;
> +
> +	default:
> +		BBDEV_LA12XX_PMD_ERR("unsupported bbdev_ipc op type");
> +		return -1;
> +	}
> +
> +	if (in_op_data->data) {
> +		data_ptr = rte_pktmbuf_mtod(in_op_data->data, char *);
> +		l1_pcie_addr = (uint32_t)GUL_USER_HUGE_PAGE_ADDR +
> +			       data_ptr - huge_start_addr;
> +		bbdev_ipc_op->in_addr = l1_pcie_addr;
> +		bbdev_ipc_op->in_len = in_op_data->length;
> +	}
> +
> +	if (out_op_data->data) {
> +		data_ptr = rte_pktmbuf_mtod(out_op_data->data, char *);
> +		l1_pcie_addr = (uint32_t)GUL_USER_HUGE_PAGE_ADDR +
> +				data_ptr - huge_start_addr;
> +		bbdev_ipc_op->out_addr = rte_cpu_to_be_32(l1_pcie_addr);
> +		bbdev_ipc_op->out_len = rte_cpu_to_be_32(out_op_data->length);
> +	}
> +
> +	/* Move Producer Index forward */
> +	pi++;
> +	/* Flip the PI flag, if wrapping */
> +	if (unlikely(q_priv->queue_size == pi)) {
> +		pi = 0;
> +		pi_flag = pi_flag ? 0 : 1;
> +	}
> +
> +	if (pi_flag)
> +		IPC_SET_PI_FLAG(pi);
> +	else
> +		IPC_RESET_PI_FLAG(pi);
> +	/* Wait for Data Copy & pi_flag update to complete before updating pi */
> +	rte_mb();
> +	/* now update pi */
> +	md->pi = rte_cpu_to_be_32(pi);
> +	q_priv->host_pi = pi;
> +
> +	BBDEV_LA12XX_PMD_DP_DEBUG(
> +			"enter: pi: %u, ci: %u, pi_flag: %u, ci_flag: %u, ring size: %u",
> +			pi, ci, pi_flag, ci_flag, q_priv->queue_size);
> +
> +	return 0;
> +}
> +
> +/* Enqueue decode burst */
> +static uint16_t
> +enqueue_dec_ops(struct rte_bbdev_queue_data *q_data,
> +		struct rte_bbdev_dec_op **ops, uint16_t nb_ops) {
> +	struct bbdev_la12xx_q_priv *q_priv = q_data->queue_private;
> +	int nb_enqueued, ret;
> +
> +	for (nb_enqueued = 0; nb_enqueued < nb_ops; nb_enqueued++) {
> +		ret = enqueue_single_op(q_priv, ops[nb_enqueued]);
> +		if (ret)
> +			break;
> +	}
> +
> +	q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
> +	q_data->queue_stats.enqueued_count += nb_enqueued;
> +
> +	return nb_enqueued;
> +}
> +
> +/* Enqueue encode burst */
> +static uint16_t
> +enqueue_enc_ops(struct rte_bbdev_queue_data *q_data,
> +		struct rte_bbdev_enc_op **ops, uint16_t nb_ops) {
> +	struct bbdev_la12xx_q_priv *q_priv = q_data->queue_private;
> +	int nb_enqueued, ret;
> +
> +	for (nb_enqueued = 0; nb_enqueued < nb_ops; nb_enqueued++) {
> +		ret = enqueue_single_op(q_priv, ops[nb_enqueued]);
> +		if (ret)
> +			break;
> +	}
> +
> +	q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
> +	q_data->queue_stats.enqueued_count += nb_enqueued;
> +
> +	return nb_enqueued;
> +}
> +
> +static inline int
> +is_bd_ring_empty(uint32_t ci, uint32_t ci_flag,
> +		 uint32_t pi, uint32_t pi_flag)
> +{
> +	if (ci == pi) {
> +		if (ci_flag == pi_flag)
> +			return 1; /* No more Buffer */
> +	}
> +	return 0;
> +}
> +
> +/* Dequeue encode burst */
> +static void *
> +dequeue_single_op(struct bbdev_la12xx_q_priv *q_priv, void *dst) {
> +	struct bbdev_la12xx_private *priv = q_priv->bbdev_priv;
> +	ipc_userspace_t *ipc_priv = priv->ipc_priv;
> +	uint32_t q_id = q_priv->q_id + HOST_RX_QUEUEID_OFFSET;
> +	ipc_instance_t *ipc_instance = ipc_priv->instance;
> +	ipc_ch_t *ch = &(ipc_instance->ch_list[q_id]);
> +	uint32_t ci, ci_flag, pi, pi_flag;
> +	ipc_br_md_t *md;
> +	void *op;
> +	uint32_t temp_pi;
> +
> +	md = &(ch->md);
> +	ci = IPC_GET_CI_INDEX(q_priv->host_ci);
> +	ci_flag = IPC_GET_CI_FLAG(q_priv->host_ci);
> +
> +	temp_pi = q_priv->host_params->pi;
> +	pi = IPC_GET_PI_INDEX(temp_pi);
> +	pi_flag = IPC_GET_PI_FLAG(temp_pi);
> +
> +	if (is_bd_ring_empty(ci, ci_flag, pi, pi_flag))
> +		return NULL;
> +
> +	BBDEV_LA12XX_PMD_DP_DEBUG(
> +		"pi: %u, ci: %u, pi_flag: %u, ci_flag: %u, ring size: %u",
> +		pi, ci, pi_flag, ci_flag, q_priv->queue_size);
> +
> +	op = q_priv->bbdev_op[ci];
> +
> +	rte_memcpy(dst, q_priv->msg_ch_vaddr[ci],
> +		sizeof(struct bbdev_ipc_enqueue_op));
> +
> +	/* Move Consumer Index forward */
> +	ci++;
> +	/* Flip the CI flag, if wrapping */
> +	if (q_priv->queue_size == ci) {
> +		ci = 0;
> +		ci_flag = ci_flag ? 0 : 1;
> +	}
> +	if (ci_flag)
> +		IPC_SET_CI_FLAG(ci);
> +	else
> +		IPC_RESET_CI_FLAG(ci);
> +	md->ci = rte_cpu_to_be_32(ci);
> +	q_priv->host_ci = ci;
> +
> +	BBDEV_LA12XX_PMD_DP_DEBUG(
> +		"exit: pi: %u, ci: %u, pi_flag: %u, ci_flag: %u, ring size: %u",
> +		pi, ci, pi_flag, ci_flag, q_priv->queue_size);
> +
>
> So you don't use any of the BBDEV flags to report CRC and syndrome parity check in the response?
that will be supported in next version.
>
> +	return op;
> +}
> +
> +/* Dequeue decode burst */
> +static uint16_t
> +dequeue_dec_ops(struct rte_bbdev_queue_data *q_data,
> +		struct rte_bbdev_dec_op **ops, uint16_t nb_ops) {
> +	struct bbdev_la12xx_q_priv *q_priv = q_data->queue_private;
> +	struct bbdev_ipc_enqueue_op bbdev_ipc_op;
> +	int nb_dequeued;
> +
> +	for (nb_dequeued = 0; nb_dequeued < nb_ops; nb_dequeued++) {
> +		ops[nb_dequeued] = dequeue_single_op(q_priv, &bbdev_ipc_op);
> +		if (!ops[nb_dequeued])
> +			break;
> +		ops[nb_dequeued]->status = bbdev_ipc_op.status;
> +	}
> +	q_data->queue_stats.dequeued_count += nb_dequeued;
> +
> +	return nb_dequeued;
> +}
> +
> +/* Dequeue encode burst */
> +static uint16_t
> +dequeue_enc_ops(struct rte_bbdev_queue_data *q_data,
> +		struct rte_bbdev_enc_op **ops, uint16_t nb_ops) {
> +	struct bbdev_la12xx_q_priv *q_priv = q_data->queue_private;
> +	struct bbdev_ipc_enqueue_op bbdev_ipc_op;
> +	int nb_enqueued;
> +
> +	for (nb_enqueued = 0; nb_enqueued < nb_ops; nb_enqueued++) {
> +		ops[nb_enqueued] = dequeue_single_op(q_priv, &bbdev_ipc_op);
> +		if (!ops[nb_enqueued])
> +			break;
> +		ops[nb_enqueued]->status = bbdev_ipc_op.status;
> +	}
> +	q_data->queue_stats.enqueued_count += nb_enqueued;
> +
> +	return nb_enqueued;
> +}
> +
>   static struct hugepage_info *
>   get_hugepage_info(void)
>   {
> @@ -720,10 +1105,14 @@ la12xx_bbdev_create(struct rte_vdev_device *vdev,
>   	bbdev->intr_handle = NULL;
>   
>   	/* register rx/tx burst functions for data path */
> -	bbdev->dequeue_enc_ops = NULL;
> -	bbdev->dequeue_dec_ops = NULL;
> -	bbdev->enqueue_enc_ops = NULL;
> -	bbdev->enqueue_dec_ops = NULL;
> +	bbdev->dequeue_enc_ops = dequeue_enc_ops;
> +	bbdev->dequeue_dec_ops = dequeue_dec_ops;
> +	bbdev->enqueue_enc_ops = enqueue_enc_ops;
> +	bbdev->enqueue_dec_ops = enqueue_dec_ops;
>
> These above are used for 4G operations, since the capability is not there thet can be null.
>
> +	bbdev->dequeue_ldpc_enc_ops = dequeue_enc_ops;
> +	bbdev->dequeue_ldpc_dec_ops = dequeue_dec_ops;
> +	bbdev->enqueue_ldpc_enc_ops = enqueue_enc_ops;
> +	bbdev->enqueue_ldpc_dec_ops = enqueue_dec_ops;
>   
>   	return 0;
>   }
> diff --git a/drivers/baseband/la12xx/bbdev_la12xx_ipc.h b/drivers/baseband/la12xx/bbdev_la12xx_ipc.h
> index 9d5789f726..4e181e9254 100644
> --- a/drivers/baseband/la12xx/bbdev_la12xx_ipc.h
> +++ b/drivers/baseband/la12xx/bbdev_la12xx_ipc.h
> @@ -76,6 +76,25 @@ typedef struct {
>   	_IOWR(GUL_IPC_MAGIC, 5, struct ipc_msg *)  #define IOCTL_GUL_IPC_CHANNEL_RAISE_INTERRUPT _IOW(GUL_IPC_MAGIC, 6, int *)
>   
> +#define GUL_USER_HUGE_PAGE_OFFSET	(0)
> +#define GUL_PCI1_ADDR_BASE	(0x00000000ULL)
> +
> +#define GUL_USER_HUGE_PAGE_ADDR	(GUL_PCI1_ADDR_BASE + GUL_USER_HUGE_PAGE_OFFSET)
> +
> +/* IPC PI/CI index & flag manipulation helpers */
> +#define IPC_PI_CI_FLAG_MASK	0x80000000 /*  (1<<31) */
> +#define IPC_PI_CI_INDEX_MASK	0x7FFFFFFF /* ~(1<<31) */
> +
> +#define IPC_SET_PI_FLAG(x)	(x |= IPC_PI_CI_FLAG_MASK)
> +#define IPC_RESET_PI_FLAG(x)	(x &= IPC_PI_CI_INDEX_MASK)
> +#define IPC_GET_PI_FLAG(x)	(x >> 31)
> +#define IPC_GET_PI_INDEX(x)	(x & IPC_PI_CI_INDEX_MASK)
> +
> +#define IPC_SET_CI_FLAG(x)	(x |= IPC_PI_CI_FLAG_MASK)
> +#define IPC_RESET_CI_FLAG(x)	(x &= IPC_PI_CI_INDEX_MASK)
> +#define IPC_GET_CI_FLAG(x)	(x >> 31)
> +#define IPC_GET_CI_INDEX(x)	(x & IPC_PI_CI_INDEX_MASK)
> +
>   /** buffer ring common metadata */
>   typedef struct ipc_bd_ring_md {
>   	volatile uint32_t pi;		/**< Producer index and flag (MSB)
> @@ -173,6 +192,24 @@ struct bbdev_ipc_enqueue_op {
>   	uint32_t rsvd;
>   };
>   
> +/** Structure specifying dequeue operation (dequeue at LA1224) */
> +struct bbdev_ipc_dequeue_op {
> +	/** Input buffer memory address */
> +	uint32_t in_addr;
> +	/** Input buffer memory length */
> +	uint32_t in_len;
> +	/** Output buffer memory address */
> +	uint32_t out_addr;
> +	/** Output buffer memory length */
> +	uint32_t out_len;
> +	/* Number of code blocks. Only set when HARQ is used */
> +	uint32_t num_code_blocks;
> +	/** Dequeue Operation flags */
> +	uint32_t op_flags;
> +	/** Shared metadata between L1 and L2 */
> +	uint32_t shared_metadata;
> +};
> +
>   /* This shared memory would be on the host side which have copy of some
>    * of the parameters which are also part of Shared BD ring. Read access
>    * of these parameters from the host side would not be over PCI.
> --
> 2.17.1
>