[dpdk-dev,v3,6/7] virtio: simple tx routine

  Changes in v3:
- Remove return at the end of void function
- Remove always_inline attribute for virtio_xmit_cleanup

bulk free of mbufs when clean used ring.
shift operation of idx could be saved if vq_free_cnt means
free slots rather than free descriptors.

TODO: rearrange vq data structure, pack the stats var together so that we could use
one vec instruction to update all of them.

Signed-off-by: Huawei Xie <huawei.xie@intel.com>
---
 drivers/net/virtio/virtio_ethdev.h      |  3 ++
 drivers/net/virtio/virtio_rxtx_simple.c | 93 +++++++++++++++++++++++++++++++++
 2 files changed, 96 insertions(+)
  

On Tue, 20 Oct 2015 23:30:06 +0800
Huawei Xie <huawei.xie@intel.com> wrote:

> +	desc_idx = (uint16_t)(vq->vq_used_cons_idx &
> +		((vq->vq_nentries >> 1) - 1));
> +	free[0] = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
> +	nb_free = 1;
> +
> +	for (i = 1; i < VIRTIO_TX_FREE_NR; i++) {
> +		m = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
> +		if (likely(m->pool == free[0]->pool))
> +			free[nb_free++] = m;
> +		else {
> +			rte_mempool_put_bulk(free[0]->pool, (void **)free,
> +				nb_free);
> +			free[0] = m;
> +			nb_free = 1;
> +		}
> +	}
> +
> +	rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);

Might be better to introduce a function in rte_mbuf.h which
does this so other drivers can use same code?

rte_pktmbuf_free_bulk(pkts[], n)
  

On 10/22/2015 10:26 AM, Jianfeng wrote: 

> -----Original Message-----
> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Huawei Xie
> Sent: Tuesday, October 20, 2015 11:30 PM
> To: dev@dpdk.org
> Subject: [dpdk-dev] [PATCH v3 6/7] virtio: simple tx routine
> 
> Changes in v3:
> - Remove return at the end of void function
> - Remove always_inline attribute for virtio_xmit_cleanup
> 
> bulk free of mbufs when clean used ring.
> shift operation of idx could be saved if vq_free_cnt means free slots rather
> than free descriptors.
> 
> TODO: rearrange vq data structure, pack the stats var together so that we
> could use one vec instruction to update all of them.
> 
> Signed-off-by: Huawei Xie <huawei.xie@intel.com>
> ---
>  drivers/net/virtio/virtio_ethdev.h      |  3 ++
>  drivers/net/virtio/virtio_rxtx_simple.c | 93
> +++++++++++++++++++++++++++++++++
>  2 files changed, 96 insertions(+)
> 
> diff --git a/drivers/net/virtio/virtio_ethdev.h
> b/drivers/net/virtio/virtio_ethdev.h
> index d7797ab..ae2d47d 100644
> --- a/drivers/net/virtio/virtio_ethdev.h
> +++ b/drivers/net/virtio/virtio_ethdev.h
> @@ -111,6 +111,9 @@ uint16_t virtio_xmit_pkts(void *tx_queue, struct
> rte_mbuf **tx_pkts,  uint16_t virtio_recv_pkts_vec(void *rx_queue, struct
> rte_mbuf **rx_pkts,
>  		uint16_t nb_pkts);
> 
> +uint16_t virtio_xmit_pkts_simple(void *tx_queue, struct rte_mbuf
> **tx_pkts,
> +		uint16_t nb_pkts);
> +
>  /*
>   * The VIRTIO_NET_F_GUEST_TSO[46] features permit the host to send us
>   * frames larger than 1514 bytes. We do not yet support software LRO diff --
> git a/drivers/net/virtio/virtio_rxtx_simple.c
> b/drivers/net/virtio/virtio_rxtx_simple.c
> index ef17562..a53d462 100644
> --- a/drivers/net/virtio/virtio_rxtx_simple.c
> +++ b/drivers/net/virtio/virtio_rxtx_simple.c
> @@ -288,6 +288,99 @@ virtio_recv_pkts_vec(void *rx_queue, struct
> rte_mbuf **rx_pkts,
>  	return nb_pkts_received;
>  }
> 
> +#define VIRTIO_TX_FREE_THRESH 32
> +#define VIRTIO_TX_MAX_FREE_BUF_SZ 32
> +#define VIRTIO_TX_FREE_NR 32
> +/* TODO: vq->tx_free_cnt could mean num of free slots so we could avoid
> +shift */ static inline void virtio_xmit_cleanup(struct virtqueue *vq) {
> +	uint16_t i, desc_idx;
> +	int nb_free = 0;
> +	struct rte_mbuf *m, *free[VIRTIO_TX_MAX_FREE_BUF_SZ];
> +
> +	desc_idx = (uint16_t)(vq->vq_used_cons_idx &
> +		((vq->vq_nentries >> 1) - 1));
> +	free[0] = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
> +	nb_free = 1;
> +
> +	for (i = 1; i < VIRTIO_TX_FREE_NR; i++) {
> +		m = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
> +		if (likely(m->pool == free[0]->pool))
> +			free[nb_free++] = m;
> +		else {
> +			rte_mempool_put_bulk(free[0]->pool, (void **)free,
> +				nb_free);
> +			free[0] = m;
> +			nb_free = 1;
> +		}
> +	}
> +
> +	rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
> +	vq->vq_used_cons_idx += VIRTIO_TX_FREE_NR;
> +	vq->vq_free_cnt += (VIRTIO_TX_FREE_NR << 1); }
> +
> +uint16_t
> +virtio_xmit_pkts_simple(void *tx_queue, struct rte_mbuf **tx_pkts,
> +	uint16_t nb_pkts)
> +{
> +	struct virtqueue *txvq = tx_queue;
> +	uint16_t nb_used;
> +	uint16_t desc_idx;
> +	struct vring_desc *start_dp;
> +	uint16_t nb_tail, nb_commit;
> +	int i;
> +	uint16_t desc_idx_max = (txvq->vq_nentries >> 1) - 1;
> +
> +	nb_used = VIRTQUEUE_NUSED(txvq);
> +	rte_compiler_barrier();
> +
> +	nb_commit = nb_pkts = RTE_MIN((txvq->vq_free_cnt >> 1),
> nb_pkts);

Here if nb_commit is zero, how about return 0 immediately?

> +	desc_idx = (uint16_t) (txvq->vq_avail_idx & desc_idx_max);
> +	start_dp = txvq->vq_ring.desc;
> +	nb_tail = (uint16_t) (desc_idx_max + 1 - desc_idx);
> +
> +	if (nb_used >= VIRTIO_TX_FREE_THRESH)
> +		virtio_xmit_cleanup(tx_queue);

If this cleanup should be put before vq_free_cnt is referenced? It's because it may free some descs to vq_free_cnt.

> +
> +	if (nb_commit >= nb_tail) {
> +		for (i = 0; i < nb_tail; i++)
> +			txvq->vq_descx[desc_idx + i].cookie = tx_pkts[i];
> +		for (i = 0; i < nb_tail; i++) {
> +			start_dp[desc_idx].addr =
> +				RTE_MBUF_DATA_DMA_ADDR(*tx_pkts);
> +			start_dp[desc_idx].len = (*tx_pkts)->pkt_len;
> +			tx_pkts++;
> +			desc_idx++;
> +		}
> +		nb_commit -= nb_tail;
> +		desc_idx = 0;
> +	}
> +	for (i = 0; i < nb_commit; i++)
> +		txvq->vq_descx[desc_idx + i].cookie = tx_pkts[i];
> +	for (i = 0; i < nb_commit; i++) {
> +		start_dp[desc_idx].addr =
> RTE_MBUF_DATA_DMA_ADDR(*tx_pkts);
> +		start_dp[desc_idx].len = (*tx_pkts)->pkt_len;
> +		tx_pkts++;
> +		desc_idx++;
> +	}
> +
> +	rte_compiler_barrier();
> +
> +	txvq->vq_free_cnt -= (uint16_t)(nb_pkts << 1);
> +	txvq->vq_avail_idx += nb_pkts;
> +	txvq->vq_ring.avail->idx = txvq->vq_avail_idx;
> +	txvq->packets += nb_pkts;
> +
> +	if (likely(nb_pkts)) {
> +		if (unlikely(virtqueue_kick_prepare(txvq)))
> +			virtqueue_notify(txvq);
> +	}
> +
> +	return nb_pkts;
> +}
> +
>  int __attribute__((cold))
>  virtio_rxq_vec_setup(struct virtqueue *rxq)  {
> --
> 1.8.1.4
  

On 10/21/2015 2:58 AM, Stephen Hemminger wrote:
> On Tue, 20 Oct 2015 23:30:06 +0800
> Huawei Xie <huawei.xie@intel.com> wrote:
>
>> +	desc_idx = (uint16_t)(vq->vq_used_cons_idx &
>> +		((vq->vq_nentries >> 1) - 1));
>> +	free[0] = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
>> +	nb_free = 1;
>> +
>> +	for (i = 1; i < VIRTIO_TX_FREE_NR; i++) {
>> +		m = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
>> +		if (likely(m->pool == free[0]->pool))
>> +			free[nb_free++] = m;
>> +		else {
>> +			rte_mempool_put_bulk(free[0]->pool, (void **)free,
>> +				nb_free);
>> +			free[0] = m;
>> +			nb_free = 1;
>> +		}
>> +	}
>> +
>> +	rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
> Might be better to introduce a function in rte_mbuf.h which
> does this so other drivers can use same code?
>
> rte_pktmbuf_free_bulk(pkts[], n)
Agree. It would be good to have a generic rte_pktmbuf_free(/alloc)_bulk.
Several other drivers and future vhost patches also use the same logic.
I prefer to implement this later as this is API change.