[v3] vhost: allocate and free packets in bulk

  Move allocation out further and perform all allocation in bulk. The same
goes for freeing packets. In the process, also rename
virtio_dev_pktmbuf_alloc to virtio_dev_pktmbuf_prep. This
function now receives an already allocated mbuf pointer.

Signed-off-by: Balazs Nemeth <bnemeth@redhat.com>
---
 lib/librte_vhost/virtio_net.c | 58 +++++++++++++++++++++++------------
 1 file changed, 38 insertions(+), 20 deletions(-)
  

On 4/16/21 10:18 AM, Balazs Nemeth wrote:
> Move allocation out further and perform all allocation in bulk. The same
> goes for freeing packets. In the process, also rename
> virtio_dev_pktmbuf_alloc to virtio_dev_pktmbuf_prep. This
> function now receives an already allocated mbuf pointer.
> 
> Signed-off-by: Balazs Nemeth <bnemeth@redhat.com>
> ---
>  lib/librte_vhost/virtio_net.c | 58 +++++++++++++++++++++++------------
>  1 file changed, 38 insertions(+), 20 deletions(-)
> 


Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>

Thanks,
Maxime
  

On Fri, Apr 16, 2021 at 10:18 AM Balazs Nemeth <bnemeth@redhat.com> wrote:
>
> Move allocation out further and perform all allocation in bulk. The same
> goes for freeing packets. In the process, also rename
> virtio_dev_pktmbuf_alloc to virtio_dev_pktmbuf_prep. This
> function now receives an already allocated mbuf pointer.
>
> Signed-off-by: Balazs Nemeth <bnemeth@redhat.com>

The title should indicate we are only touching the tx packed path.

What about tx split?
If it is too complex to rework, this can wait.


> ---
>  lib/librte_vhost/virtio_net.c | 58 +++++++++++++++++++++++------------
>  1 file changed, 38 insertions(+), 20 deletions(-)
>
> diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
> index ff39878609..d6d5636e0f 100644
> --- a/lib/librte_vhost/virtio_net.c
> +++ b/lib/librte_vhost/virtio_net.c
> @@ -2168,6 +2168,24 @@ virtio_dev_pktmbuf_alloc(struct virtio_net *dev, struct rte_mempool *mp,
>         return NULL;
>  }
>
> +static __rte_always_inline int
> +virtio_dev_pktmbuf_prep(struct virtio_net *dev, struct rte_mbuf *pkt,
> +                        uint32_t data_len)
> +{
> +       if (rte_pktmbuf_tailroom(pkt) >= data_len)
> +               return 0;
> +
> +       /* attach an external buffer if supported */
> +       if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
> +               return 0;
> +
> +       /* check if chained buffers are allowed */
> +       if (!dev->linearbuf)
> +               return 0;
> +
> +       return -1;
> +}
> +

If virtio_dev_pktmbuf_alloc() uses this new helper, we avoid
duplicating the logic.


>  static __rte_noinline uint16_t
>  virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
>         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)

[snip]

> @@ -2429,7 +2440,7 @@ static __rte_always_inline int
>  virtio_dev_tx_single_packed(struct virtio_net *dev,
>                             struct vhost_virtqueue *vq,
>                             struct rte_mempool *mbuf_pool,
> -                           struct rte_mbuf **pkts)
> +                           struct rte_mbuf *pkts)
>  {
>
>         uint16_t buf_id, desc_count = 0;
> @@ -2462,26 +2473,33 @@ virtio_dev_tx_packed(struct virtio_net *dev,
>         uint32_t pkt_idx = 0;
>         uint32_t remained = count;
>
> +       if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
> +               return 0;
> +
>         do {
>                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
>
>                 if (remained >= PACKED_BATCH_SIZE) {
> -                       if (!virtio_dev_tx_batch_packed(dev, vq, mbuf_pool,
> +                       if (!virtio_dev_tx_batch_packed(dev, vq,
>                                                         &pkts[pkt_idx])) {
>                                 pkt_idx += PACKED_BATCH_SIZE;
>                                 remained -= PACKED_BATCH_SIZE;
> +

No need for the extra line.


>                                 continue;
>                         }
>                 }
>
>                 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
> -                                               &pkts[pkt_idx]))
> +                                               pkts[pkt_idx]))
>                         break;
>                 pkt_idx++;
>                 remained--;
>
>         } while (remained);
>
> +       if (pkt_idx != count)
> +               rte_pktmbuf_free_bulk(&pkts[pkt_idx], count - pkt_idx);
> +
>         if (vq->shadow_used_idx) {
>                 do_data_copy_dequeue(vq);
>

With those comments addressed,
Reviewed-by: David Marchand <david.marchand@redhat.com>

Thanks Balazs!
  

Hi David,

I was also thinking about using the same idea for tx split so that
virtio_dev_pktmbuf_alloc can be removed completely. I don't have those
patches yet. However, I can make virtio_dev_pktmbuf_alloc use
virtio_dev_pktmbuf_prep for this patch that addresses the packed
version and submit other patches for tx split later. Would that work for
now?

Regards,
Balazs

On Fri, Apr 16, 2021 at 11:05 AM David Marchand <david.marchand@redhat.com>
wrote:

> On Fri, Apr 16, 2021 at 10:18 AM Balazs Nemeth <bnemeth@redhat.com> wrote:
> >
> > Move allocation out further and perform all allocation in bulk. The same
> > goes for freeing packets. In the process, also rename
> > virtio_dev_pktmbuf_alloc to virtio_dev_pktmbuf_prep. This
> > function now receives an already allocated mbuf pointer.
> >
> > Signed-off-by: Balazs Nemeth <bnemeth@redhat.com>
>
> The title should indicate we are only touching the tx packed path.
>
> What about tx split?
> If it is too complex to rework, this can wait.
>
>
> > ---
> >  lib/librte_vhost/virtio_net.c | 58 +++++++++++++++++++++++------------
> >  1 file changed, 38 insertions(+), 20 deletions(-)
> >
> > diff --git a/lib/librte_vhost/virtio_net.c
> b/lib/librte_vhost/virtio_net.c
> > index ff39878609..d6d5636e0f 100644
> > --- a/lib/librte_vhost/virtio_net.c
> > +++ b/lib/librte_vhost/virtio_net.c
> > @@ -2168,6 +2168,24 @@ virtio_dev_pktmbuf_alloc(struct virtio_net *dev,
> struct rte_mempool *mp,
> >         return NULL;
> >  }
> >
> > +static __rte_always_inline int
> > +virtio_dev_pktmbuf_prep(struct virtio_net *dev, struct rte_mbuf *pkt,
> > +                        uint32_t data_len)
> > +{
> > +       if (rte_pktmbuf_tailroom(pkt) >= data_len)
> > +               return 0;
> > +
> > +       /* attach an external buffer if supported */
> > +       if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
> > +               return 0;
> > +
> > +       /* check if chained buffers are allowed */
> > +       if (!dev->linearbuf)
> > +               return 0;
> > +
> > +       return -1;
> > +}
> > +
>
> If virtio_dev_pktmbuf_alloc() uses this new helper, we avoid
> duplicating the logic.
>
>
> >  static __rte_noinline uint16_t
> >  virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
> >         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t
> count)
>
> [snip]
>
> > @@ -2429,7 +2440,7 @@ static __rte_always_inline int
> >  virtio_dev_tx_single_packed(struct virtio_net *dev,
> >                             struct vhost_virtqueue *vq,
> >                             struct rte_mempool *mbuf_pool,
> > -                           struct rte_mbuf **pkts)
> > +                           struct rte_mbuf *pkts)
> >  {
> >
> >         uint16_t buf_id, desc_count = 0;
> > @@ -2462,26 +2473,33 @@ virtio_dev_tx_packed(struct virtio_net *dev,
> >         uint32_t pkt_idx = 0;
> >         uint32_t remained = count;
> >
> > +       if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
> > +               return 0;
> > +
> >         do {
> >                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
> >
> >                 if (remained >= PACKED_BATCH_SIZE) {
> > -                       if (!virtio_dev_tx_batch_packed(dev, vq,
> mbuf_pool,
> > +                       if (!virtio_dev_tx_batch_packed(dev, vq,
> >                                                         &pkts[pkt_idx]))
> {
> >                                 pkt_idx += PACKED_BATCH_SIZE;
> >                                 remained -= PACKED_BATCH_SIZE;
> > +
>
> No need for the extra line.
>
>
> >                                 continue;
> >                         }
> >                 }
> >
> >                 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
> > -                                               &pkts[pkt_idx]))
> > +                                               pkts[pkt_idx]))
> >                         break;
> >                 pkt_idx++;
> >                 remained--;
> >
> >         } while (remained);
> >
> > +       if (pkt_idx != count)
> > +               rte_pktmbuf_free_bulk(&pkts[pkt_idx], count - pkt_idx);
> > +
> >         if (vq->shadow_used_idx) {
> >                 do_data_copy_dequeue(vq);
> >
>
> With those comments addressed,
> Reviewed-by: David Marchand <david.marchand@redhat.com>
>
> Thanks Balazs!
>
>
> --
> David Marchand
>
>
  

Hi Balazs,

On 4/16/21 11:12 AM, Balazs Nemeth wrote:
> Hi David,
> 
> I was also thinking about using the same idea for tx split so that
> virtio_dev_pktmbuf_alloc can be removed completely. I don't have those
> patches yet. However, I can make virtio_dev_pktmbuf_alloc use
> virtio_dev_pktmbuf_prep for this patch that addresses the packed
> version and submit other patches for tx split later. Would that work for
> now?

I think so.
That would be great to have the same for split ring, but more rework may
be necessary than for packed ring. So I'm fine if the split ring part is
done after this release.

Thanks,
Maxime
> Regards,
> Balazs 
> 
> On Fri, Apr 16, 2021 at 11:05 AM David Marchand
> <david.marchand@redhat.com <mailto:david.marchand@redhat.com>> wrote:
> 
>     On Fri, Apr 16, 2021 at 10:18 AM Balazs Nemeth <bnemeth@redhat.com
>     <mailto:bnemeth@redhat.com>> wrote:
>     >
>     > Move allocation out further and perform all allocation in bulk.
>     The same
>     > goes for freeing packets. In the process, also rename
>     > virtio_dev_pktmbuf_alloc to virtio_dev_pktmbuf_prep. This
>     > function now receives an already allocated mbuf pointer.
>     >
>     > Signed-off-by: Balazs Nemeth <bnemeth@redhat.com
>     <mailto:bnemeth@redhat.com>>
> 
>     The title should indicate we are only touching the tx packed path.
> 
>     What about tx split?
>     If it is too complex to rework, this can wait.
> 
> 
>     > ---
>     >  lib/librte_vhost/virtio_net.c | 58
>     +++++++++++++++++++++++------------
>     >  1 file changed, 38 insertions(+), 20 deletions(-)
>     >
>     > diff --git a/lib/librte_vhost/virtio_net.c
>     b/lib/librte_vhost/virtio_net.c
>     > index ff39878609..d6d5636e0f 100644
>     > --- a/lib/librte_vhost/virtio_net.c
>     > +++ b/lib/librte_vhost/virtio_net.c
>     > @@ -2168,6 +2168,24 @@ virtio_dev_pktmbuf_alloc(struct virtio_net
>     *dev, struct rte_mempool *mp,
>     >         return NULL;
>     >  }
>     >
>     > +static __rte_always_inline int
>     > +virtio_dev_pktmbuf_prep(struct virtio_net *dev, struct rte_mbuf *pkt,
>     > +                        uint32_t data_len)
>     > +{
>     > +       if (rte_pktmbuf_tailroom(pkt) >= data_len)
>     > +               return 0;
>     > +
>     > +       /* attach an external buffer if supported */
>     > +       if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
>     > +               return 0;
>     > +
>     > +       /* check if chained buffers are allowed */
>     > +       if (!dev->linearbuf)
>     > +               return 0;
>     > +
>     > +       return -1;
>     > +}
>     > +
> 
>     If virtio_dev_pktmbuf_alloc() uses this new helper, we avoid
>     duplicating the logic.
> 
> 
>     >  static __rte_noinline uint16_t
>     >  virtio_dev_tx_split(struct virtio_net *dev, struct
>     vhost_virtqueue *vq,
>     >         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts,
>     uint16_t count)
> 
>     [snip]
> 
>     > @@ -2429,7 +2440,7 @@ static __rte_always_inline int
>     >  virtio_dev_tx_single_packed(struct virtio_net *dev,
>     >                             struct vhost_virtqueue *vq,
>     >                             struct rte_mempool *mbuf_pool,
>     > -                           struct rte_mbuf **pkts)
>     > +                           struct rte_mbuf *pkts)
>     >  {
>     >
>     >         uint16_t buf_id, desc_count = 0;
>     > @@ -2462,26 +2473,33 @@ virtio_dev_tx_packed(struct virtio_net *dev,
>     >         uint32_t pkt_idx = 0;
>     >         uint32_t remained = count;
>     >
>     > +       if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
>     > +               return 0;
>     > +
>     >         do {
>     >                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
>     >
>     >                 if (remained >= PACKED_BATCH_SIZE) {
>     > -                       if (!virtio_dev_tx_batch_packed(dev, vq,
>     mbuf_pool,
>     > +                       if (!virtio_dev_tx_batch_packed(dev, vq,
>     >                                                       
>      &pkts[pkt_idx])) {
>     >                                 pkt_idx += PACKED_BATCH_SIZE;
>     >                                 remained -= PACKED_BATCH_SIZE;
>     > +
> 
>     No need for the extra line.
> 
> 
>     >                                 continue;
>     >                         }
>     >                 }
>     >
>     >                 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
>     > -                                               &pkts[pkt_idx]))
>     > +                                               pkts[pkt_idx]))
>     >                         break;
>     >                 pkt_idx++;
>     >                 remained--;
>     >
>     >         } while (remained);
>     >
>     > +       if (pkt_idx != count)
>     > +               rte_pktmbuf_free_bulk(&pkts[pkt_idx], count -
>     pkt_idx);
>     > +
>     >         if (vq->shadow_used_idx) {
>     >                 do_data_copy_dequeue(vq);
>     >
> 
>     With those comments addressed,
>     Reviewed-by: David Marchand <david.marchand@redhat.com
>     <mailto:david.marchand@redhat.com>>
> 
>     Thanks Balazs!
> 
> 
>     -- 
>     David Marchand
>
  

On Fri, Apr 16, 2021 at 11:12 AM Balazs Nemeth <bnemeth@redhat.com> wrote:
> I was also thinking about using the same idea for tx split so that virtio_dev_pktmbuf_alloc can be removed completely. I don't have those patches yet. However, I can make virtio_dev_pktmbuf_alloc use virtio_dev_pktmbuf_prep for this patch that addresses the packed version and submit other patches for tx split later. Would that work for now?

Ok for me.

And revisit the tx split part in 21.08 :-).