Re: [dpdk-dev] [PATCH v3 5/6] raw/octeontx2_ep: add dequeue operation

[Gavin Hu <Gavin.Hu@arm.com>]

Hi Mahipal, Jerin,

> -----Original Message-----
> From: Mahipal Challa <mchalla@marvell.com>
> Sent: Monday, January 6, 2020 8:07 PM
> To: dev@dpdk.org
> Cc: jerinj@marvell.com; pathreya@marvell.com; snilla@marvell.com;
> venkatn@marvell.com; Gavin Hu <Gavin.Hu@arm.com>
> Subject: [dpdk-dev] [PATCH v3 5/6] raw/octeontx2_ep: add dequeue
> operation
> 
> Add rawdev dequeue operation for SDP VF devices.
> 
> Signed-off-by: Mahipal Challa <mchalla@marvell.com>
> ---
>  drivers/raw/octeontx2_ep/otx2_ep_enqdeq.c | 197
> ++++++++++++++++++++++++++++++
>  drivers/raw/octeontx2_ep/otx2_ep_enqdeq.h |   2 +
>  drivers/raw/octeontx2_ep/otx2_ep_rawdev.c |   1 +
>  drivers/raw/octeontx2_ep/otx2_ep_rawdev.h |  18 ++-
>  4 files changed, 217 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/raw/octeontx2_ep/otx2_ep_enqdeq.c
> b/drivers/raw/octeontx2_ep/otx2_ep_enqdeq.c
> index 6910f08..dd270b5 100644
> --- a/drivers/raw/octeontx2_ep/otx2_ep_enqdeq.c
> +++ b/drivers/raw/octeontx2_ep/otx2_ep_enqdeq.c
> @@ -260,6 +260,7 @@
>  		rte_mempool_get(sdpvf->enqdeq_mpool, &buf);
>  		if (buf == NULL) {
>  			otx2_err("OQ buffer alloc failed");
> +			droq->stats.rx_alloc_failure++;
>  			/* sdp_droq_destroy_ring_buffers(droq);*/
>  			return -ENOMEM;
>  		}
> @@ -645,3 +646,199 @@
>  	return SDP_IQ_SEND_FAILED;
>  }
> 
> +static uint32_t
> +sdp_droq_refill(struct sdp_device *sdpvf, struct sdp_droq *droq)
> +{
> +	struct sdp_droq_desc *desc_ring;
> +	uint32_t desc_refilled = 0;
> +	void *buf = NULL;
> +
> +	desc_ring = droq->desc_ring;
> +
> +	while (droq->refill_count && (desc_refilled < droq->nb_desc)) {
> +		/* If a valid buffer exists (happens if there is no dispatch),
> +		 * reuse the buffer, else allocate.
> +		 */
> +		if (droq->recv_buf_list[droq->refill_idx].buffer != NULL)
> +			break;
> +
> +		rte_mempool_get(sdpvf->enqdeq_mpool, &buf);
> +		/* If a buffer could not be allocated, no point in
> +		 * continuing
> +		 */
> +		if (buf == NULL) {
> +			droq->stats.rx_alloc_failure++;
> +			break;
> +		}
> +
> +		droq->recv_buf_list[droq->refill_idx].buffer = buf;
> +		desc_ring[droq->refill_idx].buffer_ptr =
> rte_mem_virt2iova(buf);
> +
> +		/* Reset any previous values in the length field. */
> +		droq->info_list[droq->refill_idx].length = 0;
> +
> +		droq->refill_idx = sdp_incr_index(droq->refill_idx, 1,
> +				droq->nb_desc);
> +
> +		desc_refilled++;
> +		droq->refill_count--;
> +
> +	}
> +
> +	return desc_refilled;
> +}
> +
> +static int
> +sdp_droq_read_packet(struct sdp_device *sdpvf __rte_unused,
> +		     struct sdp_droq *droq,
> +		     struct sdp_droq_pkt *droq_pkt)
> +{
> +	struct sdp_droq_info *info;
> +	uint32_t total_len = 0;
> +	uint32_t pkt_len = 0;
> +
> +	info = &droq->info_list[droq->read_idx];
> +	sdp_swap_8B_data((uint64_t *)&info->length, 1);
> +	if (!info->length) {
> +		otx2_err("OQ info_list->length[%ld]", (long)info->length);
> +		goto oq_read_fail;
> +	}
> +
> +	/* Deduce the actual data size */
> +	info->length -= SDP_RH_SIZE;
> +	total_len += (uint32_t)info->length;
> +
> +	otx2_sdp_dbg("OQ: pkt_len[%ld], buffer_size %d",
> +			(long)info->length, droq->buffer_size);
> +	if (info->length > droq->buffer_size) {
> +		otx2_err("This mode is not supported: pkt_len >
> buffer_size");
> +		goto oq_read_fail;
> +	}
> +
> +	if (info->length <= droq->buffer_size) {
> +		pkt_len = (uint32_t)info->length;
> +		droq_pkt->data = droq->recv_buf_list[droq-
> >read_idx].buffer;
> +		droq_pkt->len  = pkt_len;
> +
> +		droq->recv_buf_list[droq->read_idx].buffer = NULL;
> +		droq->read_idx = sdp_incr_index(droq->read_idx,	1,/*
> count */
> +						droq->nb_desc /* max rd
> idx */);
> +		droq->refill_count++;
> +
> +	}
> +
> +	info->length = 0;
> +
> +	return SDP_OQ_RECV_SUCCESS;
> +
> +oq_read_fail:
> +	return SDP_OQ_RECV_FAILED;
> +}
> +
> +static inline uint32_t
> +sdp_check_droq_pkts(struct sdp_droq *droq, uint32_t burst_size)
> +{
> +	uint32_t min_pkts = 0;
> +	uint32_t new_pkts;
> +	uint32_t pkt_count;
> +
> +	/* Latest available OQ packets */
> +	pkt_count = rte_read32(droq->pkts_sent_reg);
> +
> +	/* Newly arrived packets */
> +	new_pkts = pkt_count - droq->last_pkt_count;
> +	otx2_sdp_dbg("Recvd [%d] new OQ pkts", new_pkts);
> +
> +	min_pkts = (new_pkts > burst_size) ? burst_size : new_pkts;
> +	if (min_pkts) {
> +		rte_atomic64_add(&droq->pkts_pending, min_pkts);
> +		/* Back up the aggregated packet count so far */
> +		droq->last_pkt_count += min_pkts;
> +	}
> +
> +	return min_pkts;
> +}
> +
> +/* Check for response arrival from OCTEON TX2
> + * returns number of requests completed
> + */
> +int
> +sdp_rawdev_dequeue(struct rte_rawdev *rawdev,
> +		   struct rte_rawdev_buf **buffers, unsigned int count,
> +		   rte_rawdev_obj_t context __rte_unused)
> +{
> +	struct sdp_droq_pkt *oq_pkt;
> +	struct sdp_device *sdpvf;
> +	struct sdp_droq *droq;
> +
> +	uint32_t q_no = 0, pkts;
> +	uint32_t new_pkts;
> +	uint32_t ret;
> +
> +	sdpvf = (struct sdp_device *)rawdev->dev_private;
> +
> +	droq = sdpvf->droq[q_no];
> +	if (!droq) {
> +		otx2_err("Invalid droq[%d]", q_no);
> +		goto deq_fail;
> +	}
> +
> +	/* Grab the lock */
> +	rte_spinlock_lock(&droq->lock);
> +
> +	new_pkts = sdp_check_droq_pkts(droq, count);
> +	if (!new_pkts) {
> +		otx2_sdp_dbg("Zero new_pkts:%d", new_pkts);
> +		goto deq_fail; /* No pkts at this moment */
> +	}
> +
> +	otx2_sdp_dbg("Received new_pkts = %d", new_pkts);
> +
> +	for (pkts = 0; pkts < new_pkts; pkts++) {
> +
> +		/* Push the received pkt to application */
> +		oq_pkt = (struct sdp_droq_pkt *)buffers[pkts];
> +
> +		ret = sdp_droq_read_packet(sdpvf, droq, oq_pkt);
> +		if (ret) {
> +			otx2_err("DROQ read pakt failed.");
> +			goto deq_fail;
> +		}
> +
> +		/* Stats */
> +		droq->stats.pkts_received++;
> +		droq->stats.bytes_received += oq_pkt->len;
> +	}
> +
> +	/* Ack the h/w with no# of pkts read by Host */
> +	rte_write32(pkts, droq->pkts_sent_reg);
Rte_write32 has a built-in rte_io_wmb, aka "dsb" for aarch64, inside, this hurts performance, this is on the fast data path? 
> +	rte_cio_wmb();
This barrier should be hoisted up to between the dequeue operation and register write to keep the ordering of two.
I am surprised why rte_cio_mb is not defined within DPDK, it is required here to keep the read/write ordering, neither rmb nor wmb can fulfill the ordering required.
> +
> +	droq->last_pkt_count -= pkts;
> +
> +	otx2_sdp_dbg("DROQ pkts[%d] pushed to application", pkts);
> +
> +	/* Refill DROQ buffers */
> +	if (droq->refill_count >= 2 /* droq->refill_threshold */) {
> +		int desc_refilled = sdp_droq_refill(sdpvf, droq);
> +
> +		/* Flush the droq descriptor data to memory to be sure
> +		 * that when we update the credits the data in memory is
> +		 * accurate.
> +		 */
> +		rte_write32(desc_refilled, droq->pkts_credit_reg);
Rte_cio_wmb should be used instead of the built-in rte_io_wmb to improve the performance.
But if performance is not a concern here, rte_io_wmb rescued the situation. 
> +
> +		/* Ensure mmio write completes */
> +		rte_wmb();
> +		otx2_sdp_dbg("Refilled count = %d", desc_refilled);
> +	}
> +
> +	/* Release the spin lock */
> +	rte_spinlock_unlock(&droq->lock);
> +
> +	return pkts;
> +
> +deq_fail:
> +	rte_spinlock_unlock(&droq->lock);
> +	return SDP_OQ_RECV_FAILED;
> +}
> diff --git a/drivers/raw/octeontx2_ep/otx2_ep_enqdeq.h
> b/drivers/raw/octeontx2_ep/otx2_ep_enqdeq.h
> index b9b7c0b..172fdc5 100644
> --- a/drivers/raw/octeontx2_ep/otx2_ep_enqdeq.h
> +++ b/drivers/raw/octeontx2_ep/otx2_ep_enqdeq.h
> @@ -11,6 +11,8 @@
>  #define SDP_IQ_SEND_FAILED      (-1)
>  #define SDP_IQ_SEND_SUCCESS     (0)
> 
> +#define SDP_OQ_RECV_FAILED      (-1)
> +#define SDP_OQ_RECV_SUCCESS     (0)
> 
>  static inline uint64_t
>  sdp_endian_swap_8B(uint64_t _d)
> diff --git a/drivers/raw/octeontx2_ep/otx2_ep_rawdev.c
> b/drivers/raw/octeontx2_ep/otx2_ep_rawdev.c
> index 22a6beb..7158b97 100644
> --- a/drivers/raw/octeontx2_ep/otx2_ep_rawdev.c
> +++ b/drivers/raw/octeontx2_ep/otx2_ep_rawdev.c
> @@ -252,6 +252,7 @@
>  	.dev_stop       = sdp_rawdev_stop,
>  	.dev_close      = sdp_rawdev_close,
>  	.enqueue_bufs   = sdp_rawdev_enqueue,
> +	.dequeue_bufs   = sdp_rawdev_dequeue,
>  };
> 
>  static int
> diff --git a/drivers/raw/octeontx2_ep/otx2_ep_rawdev.h
> b/drivers/raw/octeontx2_ep/otx2_ep_rawdev.h
> index 8fd06fb..a77cbab 100644
> --- a/drivers/raw/octeontx2_ep/otx2_ep_rawdev.h
> +++ b/drivers/raw/octeontx2_ep/otx2_ep_rawdev.h
> @@ -279,6 +279,18 @@ struct sdp_recv_buffer {
>  };
>  #define SDP_DROQ_RECVBUF_SIZE	(sizeof(struct sdp_recv_buffer))
> 
> +/* DROQ statistics. Each output queue has four stats fields. */
> +struct sdp_droq_stats {
> +	/* Number of packets received in this queue. */
> +	uint64_t pkts_received;
> +
> +	/* Bytes received by this queue. */
> +	uint64_t bytes_received;
> +
> +	/* Num of failures of rte_pktmbuf_alloc() */
> +	uint64_t rx_alloc_failure;
> +};
> +
>  /* Structure to define the configuration attributes for each Output queue.
> */
>  struct sdp_oq_config {
>  	/* Max number of OQs available */
> @@ -345,6 +357,9 @@ struct sdp_droq {
>  	 */
>  	void *pkts_sent_reg;
> 
> +	/* Statistics for this DROQ. */
> +	struct sdp_droq_stats stats;
> +
>  	/* DMA mapped address of the DROQ descriptor ring. */
>  	size_t desc_ring_dma;
> 
> @@ -476,6 +491,7 @@ struct sdp_device {
> 
>  int sdp_rawdev_enqueue(struct rte_rawdev *dev, struct rte_rawdev_buf
> **buffers,
>  		       unsigned int count, rte_rawdev_obj_t context);
> -
> +int sdp_rawdev_dequeue(struct rte_rawdev *dev, struct rte_rawdev_buf
> **buffers,
> +		       unsigned int count, rte_rawdev_obj_t context);
> 
>  #endif /* _OTX2_EP_RAWDEV_H_ */
> --
> 1.8.3.1