[dpdk-dev] [PATCH v8 2/4] vhost: add support for packed ring in async vhost

Jiang, Cheng1 cheng1.jiang at intel.com
Tue Apr 27 08:07:00 CEST 2021


Hi Jiayu,

> -----Original Message-----
> From: Hu, Jiayu <jiayu.hu at intel.com>
> Sent: Tuesday, April 27, 2021 1:16 PM
> To: Jiang, Cheng1 <cheng1.jiang at intel.com>; maxime.coquelin at redhat.com;
> Xia, Chenbo <chenbo.xia at intel.com>
> Cc: dev at dpdk.org; Yang, YvonneX <yvonnex.yang at intel.com>; Wang, Yinan
> <yinan.wang at intel.com>; Liu, Yong <yong.liu at intel.com>
> Subject: RE: [PATCH v8 2/4] vhost: add support for packed ring in async vhost
> 
> Hi Cheng,
> 
> Some comments are inline.
> 
> > -----Original Message-----
> > From: Jiang, Cheng1 <cheng1.jiang at intel.com>
> > Sent: Monday, April 19, 2021 4:51 PM
> > To: maxime.coquelin at redhat.com; Xia, Chenbo <chenbo.xia at intel.com>
> > Cc: dev at dpdk.org; Hu, Jiayu <jiayu.hu at intel.com>; Yang, YvonneX
> > <yvonnex.yang at intel.com>; Wang, Yinan <yinan.wang at intel.com>; Liu,
> > Yong <yong.liu at intel.com>; Jiang, Cheng1 <cheng1.jiang at intel.com>
> > Subject: [PATCH v8 2/4] vhost: add support for packed ring in async
> > vhost
> >
> > For now async vhost data path only supports split ring. This patch
> > enables packed ring in async vhost data path to make async vhost
> > compatible with virtio 1.1 spec.
> >
> > Signed-off-by: Cheng Jiang <Cheng1.jiang at intel.com>
> > ---
> >  lib/librte_vhost/rte_vhost_async.h |   1 +
> >  lib/librte_vhost/vhost.c           |  79 ++++--
> >  lib/librte_vhost/vhost.h           |  15 +-
> >  lib/librte_vhost/virtio_net.c      | 441 +++++++++++++++++++++++++++--
> >  4 files changed, 488 insertions(+), 48 deletions(-)
> >
> > diff --git a/lib/librte_vhost/virtio_net.c
> > b/lib/librte_vhost/virtio_net.c index 438bdafd1..ce88ad3c0 100644
> > --- a/lib/librte_vhost/virtio_net.c
> > +++ b/lib/librte_vhost/virtio_net.c
> > @@ -363,14 +363,14 @@
> > vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue
> *vq,
> > }
> >
> >  static __rte_always_inline void
> > -vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
> > -   struct vhost_virtqueue *vq,
> > -   uint32_t len[],
> > -   uint16_t id[],
> > -   uint16_t count[],
> > +vhost_shadow_enqueue_packed(struct vhost_virtqueue *vq,
> > +   uint32_t *len,
> > +   uint16_t *id,
> > +   uint16_t *count,
> >     uint16_t num_buffers)
> >  {
> >  uint16_t i;
> > +
> >  for (i = 0; i < num_buffers; i++) {
> >  /* enqueue shadow flush action aligned with batch num */  if
> > (!vq->shadow_used_idx) @@ -382,6 +382,17 @@
> > vhost_shadow_enqueue_single_packed(struct
> > virtio_net *dev,
> >  vq->shadow_aligned_idx += count[i];
> >  vq->shadow_used_idx++;
> >  }
> > +}
> > +
> > +static __rte_always_inline void
> > +vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
> > +   struct vhost_virtqueue *vq,
> > +   uint32_t *len,
> > +   uint16_t *id,
> > +   uint16_t *count,
> > +   uint16_t num_buffers)
> > +{
> > +vhost_shadow_enqueue_packed(vq, len, id, count, num_buffers);
> >
> >  if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
> > do_data_copy_enqueue(dev, vq); @@ -1474,6 +1485,23 @@
> > store_dma_desc_info_split(struct vring_used_elem *s_ring, struct
> > vring_used_elem  }  }
> >
> > +static __rte_always_inline void
> > +store_dma_desc_info_packed(struct vring_used_elem_packed *s_ring,
> > +struct vring_used_elem_packed *d_ring, uint16_t ring_size, uint16_t
> > +s_idx, uint16_t d_idx, uint16_t
> > count)
> > +{
> > +uint16_t elem_size = sizeof(struct vring_used_elem_packed);
> > +
> > +if (d_idx + count <= ring_size) {
> > +rte_memcpy(d_ring + d_idx, s_ring + s_idx, count *
> > elem_size);
> > +} else {
> > +uint16_t size = ring_size - d_idx;
> > +
> > +rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
> > +rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) *
> > elem_size);
> > +}
> > +}
> > +
> >  static __rte_noinline uint32_t
> >  virtio_dev_rx_async_submit_split(struct virtio_net *dev,  struct
> > vhost_virtqueue *vq, uint16_t queue_id, @@ -1556,12 +1584,12 @@
> > virtio_dev_rx_async_submit_split(struct
> > virtio_net *dev,
> >   * descriptors.
> >   */
> >  from = vq->shadow_used_idx - num_buffers; -to = vq->async_desc_idx &
> > (vq->size - 1);
> > +to = vq->async_desc_idx_split & (vq->size - 1);
> >
> >  store_dma_desc_info_split(vq->shadow_used_split,
> >  vq->async_descs_split, vq->size, from, to, num_buffers);
> >
> > -vq->async_desc_idx += num_buffers;
> > +vq->async_desc_idx_split += num_buffers;
> >  vq->shadow_used_idx -= num_buffers;
> >  } else
> >  comp_pkts[num_done_pkts++] = pkts[pkt_idx]; @@ -1619,7 +1647,7 @@
> > virtio_dev_rx_async_submit_split(struct virtio_net *dev,  num_descs +=
> > pkts_info[slot_idx & (vq->size - 1)].descs;  slot_idx--;  }
> > -vq->async_desc_idx -= num_descs;
> > +vq->async_desc_idx_split -= num_descs;
> >  /* recover shadow used ring and available ring */
> > vq->shadow_used_idx -= (vq->last_avail_idx -
> >
> > async_pkts_log[num_async_pkts].last_avail_idx - @@ -1641,6 +1669,329
> > @@ virtio_dev_rx_async_submit_split(struct
> > virtio_net *dev,
> >  return pkt_idx;
> >  }
> >
> > +static __rte_always_inline void
> > +vhost_update_used_packed(struct vhost_virtqueue *vq, struct
> > +vring_used_elem_packed *shadow_ring, uint16_t count) { int i;
> > +uint16_t used_idx = vq->last_used_idx; uint16_t head_idx =
> > +vq->last_used_idx; uint16_t head_flags = 0;
> > +
> > +if (count == 0)
> > +return;
> > +
> > +/* Split loop in two to save memory barriers */ for (i = 0; i <
> > +count; i++) {
> > +vq->desc_packed[used_idx].id = shadow_ring[i].id;
> > +vq->desc_packed[used_idx].len = shadow_ring[i].len;
> > +
> > +used_idx += shadow_ring[i].count;
> > +if (used_idx >= vq->size)
> > +used_idx -= vq->size;
> > +}
> > +
> > +/* The ordering for storing desc flags needs to be enforced. */
> > +rte_atomic_thread_fence(__ATOMIC_RELEASE);
> > +
> > +for (i = 0; i < count; i++) {
> > +uint16_t flags;
> > +
> > +if (vq->shadow_used_packed[i].len)
> > +flags = VRING_DESC_F_WRITE;
> > +else
> > +flags = 0;
> > +
> > +if (vq->used_wrap_counter) {
> > +flags |= VRING_DESC_F_USED;
> > +flags |= VRING_DESC_F_AVAIL;
> > +} else {
> > +flags &= ~VRING_DESC_F_USED;
> > +flags &= ~VRING_DESC_F_AVAIL;
> > +}
> > +
> > +if (i > 0) {
> > +vq->desc_packed[vq->last_used_idx].flags = flags;
> > +} else {
> > +head_idx = vq->last_used_idx;
> > +head_flags = flags;
> > +}
> > +
> > +vq_inc_last_used_packed(vq, shadow_ring[i].count); }
> > +
> > +vq->desc_packed[head_idx].flags = head_flags;
> > +}
> > +
> > +static __rte_always_inline int
> > +vhost_enqueue_async_single_packed(struct virtio_net *dev,
> > +    struct vhost_virtqueue *vq,
> > +    struct rte_mbuf *pkt,
> > +    struct buf_vector *buf_vec,
> > +    uint16_t *nr_descs,
> > +    uint16_t *nr_buffers,
> > +    struct vring_packed_desc *async_descs,
> > +    struct iovec *src_iovec, struct iovec *dst_iovec,
> > +    struct rte_vhost_iov_iter *src_it,
> > +    struct rte_vhost_iov_iter *dst_it) { uint16_t nr_vec = 0;
> > +uint16_t avail_idx = vq->last_avail_idx; uint16_t max_tries, tries =
> > +0; uint16_t buf_id = 0; uint32_t len = 0; uint16_t desc_count = 0;
> > +uint32_t size = pkt->pkt_len + sizeof(struct
> > virtio_net_hdr_mrg_rxbuf);
> > +uint32_t buffer_len[vq->size];
> > +uint16_t buffer_buf_id[vq->size];
> > +uint16_t buffer_desc_count[vq->size]; *nr_buffers = 0;
> nr_buffers and nr_descs are pointers of num_buffers and num_desc in
> virtio_dev_rx_async_submit_packed(), and num_buffers and num_desc
> don't have init values. I think you need to init them before pass their pointers
> to another function, as it will read/update values pointed by the pointers.
> 
> In addition, *nr_buffers is set to 0, but *nr_descs is not, and both of them
> are set to
> 0 in virtio_dev_rx_async_single_packed(). It looks strange.

Sure, I'll init them in virtio_dev_rx_async_submit_packed().

> 
> > +
> > +if (rxvq_is_mergeable(dev))
> > +max_tries = vq->size - 1;
> > +else
> > +max_tries = 1;
> > +
> > +while (size > 0) {
> > +/*
> > + * if we tried all available ring items, and still
> > + * can't get enough buf, it means something abnormal
> > + * happened.
> > + */
> > +if (unlikely(++tries > max_tries))
> > +return -1;
> > +
> > +if (unlikely(fill_vec_buf_packed(dev, vq, avail_idx,
> > &desc_count, buf_vec, &nr_vec,
> > +&buf_id, &len,
> > VHOST_ACCESS_RW) < 0))
> > +return -1;
> > +
> > +len = RTE_MIN(len, size);
> > +size -= len;
> > +
> > +buffer_len[*nr_buffers] = len;
> > +buffer_buf_id[*nr_buffers] = buf_id;
> > +buffer_desc_count[*nr_buffers] = desc_count; *nr_buffers += 1;
> > +
> > +*nr_descs += desc_count;
> > +avail_idx += desc_count;
> > +if (avail_idx >= vq->size)
> > +avail_idx -= vq->size;
> > +}
> > +
> > +if (async_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, *nr_buffers,
> > src_iovec, dst_iovec,
> > +src_it, dst_it) < 0)
> > +return -1;
> > +/* store descriptors for DMA */
> > +if (avail_idx >= *nr_descs) {
> > +rte_memcpy(async_descs, &vq->desc_packed[vq-
> > >last_avail_idx],
> > +*nr_descs * sizeof(struct vring_packed_desc)); } else { uint16_t
> > +nr_copy = vq->size - vq->last_avail_idx;
> 
> It needs a blank line.

OK.

> 
> > +rte_memcpy(async_descs, &vq->desc_packed[vq-
> > >last_avail_idx],
> > +nr_copy * sizeof(struct vring_packed_desc)); rte_memcpy(async_descs +
> > +nr_copy, vq->desc_packed, (*nr_descs - nr_copy) * sizeof(struct
> > vring_packed_desc));
> > +}
> > +
> > +vhost_shadow_enqueue_packed(vq, buffer_len, buffer_buf_id,
> > buffer_desc_count, *nr_buffers);
> > +
> > +return 0;
> > +}
> > +
> > +static __rte_always_inline int16_t
> > +virtio_dev_rx_async_single_packed(struct virtio_net *dev, struct
> > vhost_virtqueue *vq,
> > +    struct rte_mbuf *pkt, uint16_t *nr_descs, uint16_t
> > *nr_buffers,
> > +    struct vring_packed_desc *async_descs,
> > +    struct iovec *src_iovec, struct iovec *dst_iovec,
> > +    struct rte_vhost_iov_iter *src_it, struct
> > rte_vhost_iov_iter *dst_it)
> > +{
> > +struct buf_vector buf_vec[BUF_VECTOR_MAX]; *nr_descs = 0;
> *nr_buffers
> > += 0;
> > +
> > +if (unlikely(vhost_enqueue_async_single_packed(dev, vq, pkt,
> > buf_vec, nr_descs, nr_buffers,
> > + async_descs, src_iovec,
> > dst_iovec,
> > + src_it, dst_it) < 0)) {
> > +VHOST_LOG_DATA(DEBUG, "(%d) failed to get enough desc
> > from vring\n", dev->vid);
> > +return -1;
> > +}
> > +
> > +VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end
> > index %d\n",
> > +dev->vid, vq->last_avail_idx, vq->last_avail_idx +
> > *nr_descs);
> > +
> > +return 0;
> > +}
> > +
> > +static __rte_always_inline void
> > +dma_error_handler_packed(struct vhost_virtqueue *vq, struct
> > vring_packed_desc *async_descs,
> > +uint16_t async_descs_idx, uint16_t slot_idx, uint32_t
> > nr_err,
> > +uint32_t *pkt_idx, uint32_t *num_async_pkts,
> > uint32_t *num_done_pkts)
> > +{
> > +uint16_t descs_err = 0;
> > +uint16_t buffers_err = 0;
> > +struct async_inflight_info *pkts_info = vq->async_pkts_info;
> > +
> > +*num_async_pkts -= nr_err;
> > +*pkt_idx -= nr_err;
> > +/* calculate the sum of buffers and descs of DMA-error packets. */
> > +while (nr_err-- > 0) { descs_err += pkts_info[slot_idx %
> > +vq->size].descs; buffers_err += pkts_info[slot_idx %
> > +vq->size].nr_buffers; slot_idx--; }
> > +
> > +vq->async_buffer_idx_packed -= buffers_err;
> > +
> > +if (vq->last_avail_idx >= descs_err) {
> > +vq->last_avail_idx -= descs_err;
> > +
> > +rte_memcpy(&vq->desc_packed[vq->last_avail_idx],
> > +&async_descs[async_descs_idx - descs_err], descs_err * sizeof(struct
> > +vring_packed_desc)); } else { uint16_t nr_copy;
> > +
> > +vq->last_avail_idx = vq->last_avail_idx + vq->size - descs_err;
> > +nr_copy = vq->size - vq->last_avail_idx;
> > +rte_memcpy(&vq->desc_packed[vq->last_avail_idx],
> > +&async_descs[async_descs_idx - descs_err], nr_copy * sizeof(struct
> > +vring_packed_desc)); descs_err -= nr_copy;
> > +rte_memcpy(&vq->desc_packed[0],
> > &async_descs[async_descs_idx - descs_err],
> > +descs_err * sizeof(struct vring_packed_desc));
> > +vq->avail_wrap_counter ^= 1;
> > +}
> > +
> > +*num_done_pkts = *pkt_idx - *num_async_pkts; }
> > +
> > +static __rte_noinline uint32_t
> > +virtio_dev_rx_async_submit_packed(struct virtio_net *dev, struct
> > +vhost_virtqueue *vq, uint16_t queue_id, struct rte_mbuf **pkts,
> > +uint32_t count, struct rte_mbuf **comp_pkts, uint32_t *comp_count) {
> > +uint32_t pkt_idx = 0, pkt_burst_idx = 0; uint16_t async_descs_idx =
> > +0; uint16_t num_buffers; uint16_t num_desc;
> How about using "num_descs" to make naming consist with "num_buffers"?

Sure, that make sense, thanks a lot.
Cheng

> 
> Thanks,
> Jiayu



More information about the dev mailing list