Hi Xuan,
> -----Original Message-----
> From: Ding, Xuan <xuan.ding@intel.com>
> Sent: Monday, May 16, 2022 10:43 AM
> To: maxime.coquelin@redhat.com; Xia, Chenbo <chenbo.xia@intel.com>
> Cc: dev@dpdk.org; Hu, Jiayu <jiayu.hu@intel.com>; Jiang, Cheng1
> <cheng1.jiang@intel.com>; Pai G, Sunil <sunil.pai.g@intel.com>;
> liangma@liangbit.com; Ding, Xuan <xuan.ding@intel.com>; Wang, YuanX
> <yuanx.wang@intel.com>
> Subject: [PATCH v7 4/5] vhost: support async dequeue for split ring
>
> From: Xuan Ding <xuan.ding@intel.com>
>
> This patch implements asynchronous dequeue data path for vhost split ring,
> a new API rte_vhost_async_try_dequeue_burst() is introduced.
>
> Signed-off-by: Xuan Ding <xuan.ding@intel.com>
> Signed-off-by: Yuan Wang <yuanx.wang@intel.com>
> Tested-by: Yvonne Yang <yvonnex.yang@intel.com>
> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
> ---
> doc/guides/prog_guide/vhost_lib.rst | 7 +
> doc/guides/rel_notes/release_22_07.rst | 5 +
> lib/vhost/rte_vhost_async.h | 37 +++
> lib/vhost/version.map | 2 +-
> lib/vhost/virtio_net.c | 337 +++++++++++++++++++++++++
> 5 files changed, 387 insertions(+), 1 deletion(-)
>
> diff --git a/doc/guides/prog_guide/vhost_lib.rst
> b/doc/guides/prog_guide/vhost_lib.rst
> index f287b76ebf..09c1c24b48 100644
> --- a/doc/guides/prog_guide/vhost_lib.rst
> +++ b/doc/guides/prog_guide/vhost_lib.rst
> @@ -282,6 +282,13 @@ The following is an overview of some key Vhost API
> functions:
> Clear inflight packets which are submitted to DMA engine in vhost async
> data
> path. Completed packets are returned to applications through ``pkts``.
>
> +* ``rte_vhost_async_try_dequeue_burst(int vid, uint16_t queue_id,
> + struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
> + int *nr_inflight, uint16_t dma_id, uint16_t vchan_id)``
In dmadev library, dma_id is int16_t, rather than uint16_t.
> +
> + Receives (dequeues) ``count`` packets from guest to host in async
> + data path, and stored them at ``pkts``.
> +
> Vhost-user Implementations
> --------------------------
>
> diff --git a/doc/guides/rel_notes/release_22_07.rst
> b/doc/guides/rel_notes/release_22_07.rst
> index 88b1e478d4..564d88623e 100644
> --- a/doc/guides/rel_notes/release_22_07.rst
> +++ b/doc/guides/rel_notes/release_22_07.rst
> @@ -70,6 +70,11 @@ New Features
> Added an API which can get the number of inflight packets in
> vhost async data path without using lock.
>
> +* **Added vhost async dequeue API to receive pkts from guest.**
> +
> + Added vhost async dequeue API which can leverage DMA devices to
> + accelerate receiving pkts from guest.
> +
> Removed Items
> -------------
>
> diff --git a/lib/vhost/rte_vhost_async.h b/lib/vhost/rte_vhost_async.h index
> 70234debf9..2789492e38 100644
> --- a/lib/vhost/rte_vhost_async.h
> +++ b/lib/vhost/rte_vhost_async.h
> @@ -204,6 +204,43 @@ uint16_t rte_vhost_clear_queue_thread_unsafe(int
> vid, uint16_t queue_id, __rte_experimental int
> rte_vhost_async_dma_configure(int16_t dma_id, uint16_t vchan_id);
>
> +/**
> + * @warning
> + * @b EXPERIMENTAL: this API may change, or be removed, without prior
> +notice
> + *
> + * This function tries to receive packets from the guest with
> +offloading
> + * copies to the async channel. The packets that are transfer completed
In rte_vhost_async.h, DMA vChannel is referred as async channel or async copy
engine. But I think it's better to replace all with DMA vChannel to be consistent
with DPDK.
> + * are returned in "pkts". The other packets that their copies are
> +submitted to
> + * the async channel but not completed are called "in-flight packets".
> + * This function will not return in-flight packets until their copies
> +are
> + * completed by the async channel.
> + *
> + * @param vid
> + * ID of vhost device to dequeue data
> + * @param queue_id
> + * ID of virtqueue to dequeue data
> + * @param mbuf_pool
> + * Mbuf_pool where host mbuf is allocated
> + * @param pkts
> + * Blank array to keep successfully dequeued packets
> + * @param count
> + * Size of the packet array
> + * @param nr_inflight
> + * >= 0: The amount of in-flight packets
Better to add more descriptions about the meaning of "nr_inflight".
Thanks,
Jiayu
Hi Jiayu,
> -----Original Message-----
> From: Hu, Jiayu <jiayu.hu@intel.com>
> Sent: Monday, May 16, 2022 1:53 PM
> To: Ding, Xuan <xuan.ding@intel.com>; maxime.coquelin@redhat.com; Xia,
> Chenbo <chenbo.xia@intel.com>
> Cc: dev@dpdk.org; Jiang, Cheng1 <cheng1.jiang@intel.com>; Pai G, Sunil
> <sunil.pai.g@intel.com>; liangma@liangbit.com; Wang, YuanX
> <yuanx.wang@intel.com>
> Subject: RE: [PATCH v7 4/5] vhost: support async dequeue for split ring
>
> Hi Xuan,
>
> > -----Original Message-----
> > From: Ding, Xuan <xuan.ding@intel.com>
> > Sent: Monday, May 16, 2022 10:43 AM
> > To: maxime.coquelin@redhat.com; Xia, Chenbo <chenbo.xia@intel.com>
> > Cc: dev@dpdk.org; Hu, Jiayu <jiayu.hu@intel.com>; Jiang, Cheng1
> > <cheng1.jiang@intel.com>; Pai G, Sunil <sunil.pai.g@intel.com>;
> > liangma@liangbit.com; Ding, Xuan <xuan.ding@intel.com>; Wang, YuanX
> > <yuanx.wang@intel.com>
> > Subject: [PATCH v7 4/5] vhost: support async dequeue for split ring
> >
> > From: Xuan Ding <xuan.ding@intel.com>
> >
> > This patch implements asynchronous dequeue data path for vhost split
> > ring, a new API rte_vhost_async_try_dequeue_burst() is introduced.
> >
> > Signed-off-by: Xuan Ding <xuan.ding@intel.com>
> > Signed-off-by: Yuan Wang <yuanx.wang@intel.com>
> > Tested-by: Yvonne Yang <yvonnex.yang@intel.com>
> > Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
> > ---
> > doc/guides/prog_guide/vhost_lib.rst | 7 +
> > doc/guides/rel_notes/release_22_07.rst | 5 +
> > lib/vhost/rte_vhost_async.h | 37 +++
> > lib/vhost/version.map | 2 +-
> > lib/vhost/virtio_net.c | 337 +++++++++++++++++++++++++
> > 5 files changed, 387 insertions(+), 1 deletion(-)
> >
> > diff --git a/doc/guides/prog_guide/vhost_lib.rst
> > b/doc/guides/prog_guide/vhost_lib.rst
> > index f287b76ebf..09c1c24b48 100644
> > --- a/doc/guides/prog_guide/vhost_lib.rst
> > +++ b/doc/guides/prog_guide/vhost_lib.rst
> > @@ -282,6 +282,13 @@ The following is an overview of some key Vhost
> > API
> > functions:
> > Clear inflight packets which are submitted to DMA engine in vhost
> > async data
> > path. Completed packets are returned to applications through ``pkts``.
> >
> > +* ``rte_vhost_async_try_dequeue_burst(int vid, uint16_t queue_id,
> > + struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t
> > +count, int *nr_inflight, uint16_t dma_id, uint16_t vchan_id)``
>
> In dmadev library, dma_id is int16_t, rather than uint16_t.
Thanks for the reminder. I will fix it in v8.
>
> > +
> > + Receives (dequeues) ``count`` packets from guest to host in async
> > + data path, and stored them at ``pkts``.
> > +
> > Vhost-user Implementations
> > --------------------------
> >
> > diff --git a/doc/guides/rel_notes/release_22_07.rst
> > b/doc/guides/rel_notes/release_22_07.rst
> > index 88b1e478d4..564d88623e 100644
> > --- a/doc/guides/rel_notes/release_22_07.rst
> > +++ b/doc/guides/rel_notes/release_22_07.rst
> > @@ -70,6 +70,11 @@ New Features
> > Added an API which can get the number of inflight packets in
> > vhost async data path without using lock.
> >
> > +* **Added vhost async dequeue API to receive pkts from guest.**
> > +
> > + Added vhost async dequeue API which can leverage DMA devices to
> > + accelerate receiving pkts from guest.
> > +
> > Removed Items
> > -------------
> >
> > diff --git a/lib/vhost/rte_vhost_async.h b/lib/vhost/rte_vhost_async.h
> > index
> > 70234debf9..2789492e38 100644
> > --- a/lib/vhost/rte_vhost_async.h
> > +++ b/lib/vhost/rte_vhost_async.h
> > @@ -204,6 +204,43 @@ uint16_t
> rte_vhost_clear_queue_thread_unsafe(int
> > vid, uint16_t queue_id, __rte_experimental int
> > rte_vhost_async_dma_configure(int16_t dma_id, uint16_t vchan_id);
> >
> > +/**
> > + * @warning
> > + * @b EXPERIMENTAL: this API may change, or be removed, without prior
> > +notice
> > + *
> > + * This function tries to receive packets from the guest with
> > +offloading
> > + * copies to the async channel. The packets that are transfer
> > +completed
>
> In rte_vhost_async.h, DMA vChannel is referred as async channel or async
> copy engine. But I think it's better to replace all with DMA vChannel to be
> consistent with DPDK.
Thanks for your suggestion, I will refine the API description in next version.
>
> > + * are returned in "pkts". The other packets that their copies are
> > +submitted to
> > + * the async channel but not completed are called "in-flight packets".
> > + * This function will not return in-flight packets until their copies
> > +are
> > + * completed by the async channel.
> > + *
> > + * @param vid
> > + * ID of vhost device to dequeue data
> > + * @param queue_id
> > + * ID of virtqueue to dequeue data
> > + * @param mbuf_pool
> > + * Mbuf_pool where host mbuf is allocated
> > + * @param pkts
> > + * Blank array to keep successfully dequeued packets
> > + * @param count
> > + * Size of the packet array
> > + * @param nr_inflight
> > + * >= 0: The amount of in-flight packets
>
> Better to add more descriptions about the meaning of "nr_inflight".
Same as above.
Regards,
Xuan
>
> Thanks,
> Jiayu
@@ -282,6 +282,13 @@ The following is an overview of some key Vhost API functions:
Clear inflight packets which are submitted to DMA engine in vhost async data
path. Completed packets are returned to applications through ``pkts``.
+* ``rte_vhost_async_try_dequeue_burst(int vid, uint16_t queue_id,
+ struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
+ int *nr_inflight, uint16_t dma_id, uint16_t vchan_id)``
+
+ Receives (dequeues) ``count`` packets from guest to host in async data path,
+ and stored them at ``pkts``.
+
Vhost-user Implementations
--------------------------
@@ -70,6 +70,11 @@ New Features
Added an API which can get the number of inflight packets in
vhost async data path without using lock.
+* **Added vhost async dequeue API to receive pkts from guest.**
+
+ Added vhost async dequeue API which can leverage DMA devices to
+ accelerate receiving pkts from guest.
+
Removed Items
-------------
@@ -204,6 +204,43 @@ uint16_t rte_vhost_clear_queue_thread_unsafe(int vid, uint16_t queue_id,
__rte_experimental
int rte_vhost_async_dma_configure(int16_t dma_id, uint16_t vchan_id);
+/**
+ * @warning
+ * @b EXPERIMENTAL: this API may change, or be removed, without prior notice
+ *
+ * This function tries to receive packets from the guest with offloading
+ * copies to the async channel. The packets that are transfer completed
+ * are returned in "pkts". The other packets that their copies are submitted to
+ * the async channel but not completed are called "in-flight packets".
+ * This function will not return in-flight packets until their copies are
+ * completed by the async channel.
+ *
+ * @param vid
+ * ID of vhost device to dequeue data
+ * @param queue_id
+ * ID of virtqueue to dequeue data
+ * @param mbuf_pool
+ * Mbuf_pool where host mbuf is allocated
+ * @param pkts
+ * Blank array to keep successfully dequeued packets
+ * @param count
+ * Size of the packet array
+ * @param nr_inflight
+ * >= 0: The amount of in-flight packets
+ * -1: Meaningless, indicates failed lock acquisition or invalid queue_id/dma_id
+ * @param dma_id
+ * The identifier of DMA device
+ * @param vchan_id
+ * The identifier of virtual DMA channel
+ * @return
+ * Number of successfully dequeued packets
+ */
+__rte_experimental
+uint16_t
+rte_vhost_async_try_dequeue_burst(int vid, uint16_t queue_id,
+ struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
+ int *nr_inflight, uint16_t dma_id, uint16_t vchan_id);
+
#ifdef __cplusplus
}
#endif
@@ -90,7 +90,7 @@ EXPERIMENTAL {
# added in 22.07
rte_vhost_async_get_inflight_thread_unsafe;
-
+ rte_vhost_async_try_dequeue_burst;
};
INTERNAL {
@@ -3171,3 +3171,340 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
return count;
}
+
+static __rte_always_inline uint16_t
+async_poll_dequeue_completed_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
+ struct rte_mbuf **pkts, uint16_t count, uint16_t dma_id,
+ uint16_t vchan_id, bool legacy_ol_flags)
+{
+ uint16_t start_idx, from, i;
+ uint16_t nr_cpl_pkts = 0;
+ struct async_inflight_info *pkts_info = vq->async->pkts_info;
+
+ vhost_async_dma_check_completed(dev, dma_id, vchan_id, VHOST_DMA_MAX_COPY_COMPLETE);
+
+ start_idx = async_get_first_inflight_pkt_idx(vq);
+
+ from = start_idx;
+ while (vq->async->pkts_cmpl_flag[from] && count--) {
+ vq->async->pkts_cmpl_flag[from] = false;
+ from = (from + 1) & (vq->size - 1);
+ nr_cpl_pkts++;
+ }
+
+ if (nr_cpl_pkts == 0)
+ return 0;
+
+ for (i = 0; i < nr_cpl_pkts; i++) {
+ from = (start_idx + i) & (vq->size - 1);
+ pkts[i] = pkts_info[from].mbuf;
+
+ if (virtio_net_with_host_offload(dev))
+ vhost_dequeue_offload(dev, &pkts_info[from].nethdr, pkts[i],
+ legacy_ol_flags);
+ }
+
+ /* write back completed descs to used ring and update used idx */
+ write_back_completed_descs_split(vq, nr_cpl_pkts);
+ __atomic_add_fetch(&vq->used->idx, nr_cpl_pkts, __ATOMIC_RELEASE);
+ vhost_vring_call_split(dev, vq);
+
+ vq->async->pkts_inflight_n -= nr_cpl_pkts;
+
+ return nr_cpl_pkts;
+}
+
+static __rte_always_inline uint16_t
+virtio_dev_tx_async_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
+ struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
+ uint16_t dma_id, uint16_t vchan_id, bool legacy_ol_flags)
+{
+ static bool allocerr_warned;
+ bool dropped = false;
+ uint16_t free_entries;
+ uint16_t pkt_idx, slot_idx = 0;
+ uint16_t nr_done_pkts = 0;
+ uint16_t pkt_err = 0;
+ uint16_t n_xfer;
+ struct vhost_async *async = vq->async;
+ struct async_inflight_info *pkts_info = async->pkts_info;
+ struct rte_mbuf *pkts_prealloc[MAX_PKT_BURST];
+ uint16_t pkts_size = count;
+
+ /**
+ * The ordering between avail index and
+ * desc reads needs to be enforced.
+ */
+ free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
+ vq->last_avail_idx;
+ if (free_entries == 0)
+ goto out;
+
+ rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
+
+ async_iter_reset(async);
+
+ count = RTE_MIN(count, MAX_PKT_BURST);
+ count = RTE_MIN(count, free_entries);
+ VHOST_LOG_DATA(DEBUG, "(%s) about to dequeue %u buffers\n",
+ dev->ifname, count);
+
+ if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts_prealloc, count))
+ goto out;
+
+ for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
+ uint16_t head_idx = 0;
+ uint16_t nr_vec = 0;
+ uint16_t to;
+ uint32_t buf_len;
+ int err;
+ struct buf_vector buf_vec[BUF_VECTOR_MAX];
+ struct rte_mbuf *pkt = pkts_prealloc[pkt_idx];
+
+ if (unlikely(fill_vec_buf_split(dev, vq, vq->last_avail_idx,
+ &nr_vec, buf_vec,
+ &head_idx, &buf_len,
+ VHOST_ACCESS_RO) < 0)) {
+ dropped = true;
+ break;
+ }
+
+ err = virtio_dev_pktmbuf_prep(dev, pkt, buf_len);
+ if (unlikely(err)) {
+ /**
+ * mbuf allocation fails for jumbo packets when external
+ * buffer allocation is not allowed and linear buffer
+ * is required. Drop this packet.
+ */
+ if (!allocerr_warned) {
+ VHOST_LOG_DATA(ERR,
+ "(%s) %s: Failed mbuf alloc of size %d from %s\n",
+ dev->ifname, __func__, buf_len, mbuf_pool->name);
+ allocerr_warned = true;
+ }
+ dropped = true;
+ break;
+ }
+
+ slot_idx = (async->pkts_idx + pkt_idx) & (vq->size - 1);
+ err = desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkt, mbuf_pool,
+ legacy_ol_flags, slot_idx, true);
+ if (unlikely(err)) {
+ if (!allocerr_warned) {
+ VHOST_LOG_DATA(ERR,
+ "(%s) %s: Failed to offload copies to async channel.\n",
+ dev->ifname, __func__);
+ allocerr_warned = true;
+ }
+ dropped = true;
+ break;
+ }
+
+ pkts_info[slot_idx].mbuf = pkt;
+
+ /* store used descs */
+ to = async->desc_idx_split & (vq->size - 1);
+ async->descs_split[to].id = head_idx;
+ async->descs_split[to].len = 0;
+ async->desc_idx_split++;
+
+ vq->last_avail_idx++;
+ }
+
+ if (unlikely(dropped))
+ rte_pktmbuf_free_bulk(&pkts_prealloc[pkt_idx], count - pkt_idx);
+
+ n_xfer = vhost_async_dma_transfer(dev, vq, dma_id, vchan_id, async->pkts_idx,
+ async->iov_iter, pkt_idx);
+
+ async->pkts_inflight_n += n_xfer;
+
+ pkt_err = pkt_idx - n_xfer;
+ if (unlikely(pkt_err)) {
+ VHOST_LOG_DATA(DEBUG, "(%s) %s: failed to transfer data.\n",
+ dev->ifname, __func__);
+
+ pkt_idx = n_xfer;
+ /* recover available ring */
+ vq->last_avail_idx -= pkt_err;
+
+ /**
+ * recover async channel copy related structures and free pktmbufs
+ * for error pkts.
+ */
+ async->desc_idx_split -= pkt_err;
+ while (pkt_err-- > 0) {
+ rte_pktmbuf_free(pkts_info[slot_idx & (vq->size - 1)].mbuf);
+ slot_idx--;
+ }
+ }
+
+ async->pkts_idx += pkt_idx;
+ if (async->pkts_idx >= vq->size)
+ async->pkts_idx -= vq->size;
+
+out:
+ /* DMA device may serve other queues, unconditionally check completed. */
+ nr_done_pkts = async_poll_dequeue_completed_split(dev, vq, pkts, pkts_size,
+ dma_id, vchan_id, legacy_ol_flags);
+
+ return nr_done_pkts;
+}
+
+__rte_noinline
+static uint16_t
+virtio_dev_tx_async_split_legacy(struct virtio_net *dev,
+ struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
+ struct rte_mbuf **pkts, uint16_t count,
+ uint16_t dma_id, uint16_t vchan_id)
+{
+ return virtio_dev_tx_async_split(dev, vq, mbuf_pool,
+ pkts, count, dma_id, vchan_id, true);
+}
+
+__rte_noinline
+static uint16_t
+virtio_dev_tx_async_split_compliant(struct virtio_net *dev,
+ struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
+ struct rte_mbuf **pkts, uint16_t count,
+ uint16_t dma_id, uint16_t vchan_id)
+{
+ return virtio_dev_tx_async_split(dev, vq, mbuf_pool,
+ pkts, count, dma_id, vchan_id, false);
+}
+
+uint16_t
+rte_vhost_async_try_dequeue_burst(int vid, uint16_t queue_id,
+ struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
+ int *nr_inflight, uint16_t dma_id, uint16_t vchan_id)
+{
+ struct virtio_net *dev;
+ struct rte_mbuf *rarp_mbuf = NULL;
+ struct vhost_virtqueue *vq;
+ int16_t success = 1;
+
+ dev = get_device(vid);
+ if (!dev || !nr_inflight)
+ return 0;
+
+ *nr_inflight = -1;
+
+ if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
+ VHOST_LOG_DATA(ERR, "(%s) %s: built-in vhost net backend is disabled.\n",
+ dev->ifname, __func__);
+ return 0;
+ }
+
+ if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
+ VHOST_LOG_DATA(ERR, "(%s) %s: invalid virtqueue idx %d.\n",
+ dev->ifname, __func__, queue_id);
+ return 0;
+ }
+
+ if (unlikely(dma_id >= RTE_DMADEV_DEFAULT_MAX)) {
+ VHOST_LOG_DATA(ERR, "(%s) %s: invalid dma id %u.\n",
+ dev->ifname, __func__, dma_id);
+ return 0;
+ }
+
+ if (unlikely(!dma_copy_track[dma_id].vchans ||
+ !dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr)) {
+ VHOST_LOG_DATA(ERR, "(%s) %s: invalid channel %d:%u.\n", dev->ifname, __func__,
+ dma_id, vchan_id);
+ return 0;
+ }
+
+ vq = dev->virtqueue[queue_id];
+
+ if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
+ return 0;
+
+ if (unlikely(vq->enabled == 0)) {
+ count = 0;
+ goto out_access_unlock;
+ }
+
+ if (unlikely(!vq->async)) {
+ VHOST_LOG_DATA(ERR, "(%s) %s: async not registered for queue id %d.\n",
+ dev->ifname, __func__, queue_id);
+ count = 0;
+ goto out_access_unlock;
+ }
+
+ if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
+ vhost_user_iotlb_rd_lock(vq);
+
+ if (unlikely(vq->access_ok == 0))
+ if (unlikely(vring_translate(dev, vq) < 0)) {
+ count = 0;
+ goto out;
+ }
+
+ /*
+ * Construct a RARP broadcast packet, and inject it to the "pkts"
+ * array, to looks like that guest actually send such packet.
+ *
+ * Check user_send_rarp() for more information.
+ *
+ * broadcast_rarp shares a cacheline in the virtio_net structure
+ * with some fields that are accessed during enqueue and
+ * __atomic_compare_exchange_n causes a write if performed compare
+ * and exchange. This could result in false sharing between enqueue
+ * and dequeue.
+ *
+ * Prevent unnecessary false sharing by reading broadcast_rarp first
+ * and only performing compare and exchange if the read indicates it
+ * is likely to be set.
+ */
+ if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
+ __atomic_compare_exchange_n(&dev->broadcast_rarp,
+ &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
+
+ rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
+ if (rarp_mbuf == NULL) {
+ VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
+ count = 0;
+ goto out;
+ }
+ /*
+ * Inject it to the head of "pkts" array, so that switch's mac
+ * learning table will get updated first.
+ */
+ pkts[0] = rarp_mbuf;
+ pkts++;
+ count -= 1;
+ }
+
+ if (unlikely(vq_is_packed(dev))) {
+ static bool not_support_pack_log;
+ if (!not_support_pack_log) {
+ VHOST_LOG_DATA(ERR,
+ "(%s) %s: async dequeue does not support packed ring.\n",
+ dev->ifname, __func__);
+ not_support_pack_log = true;
+ }
+ count = 0;
+ goto out;
+ }
+
+ if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
+ count = virtio_dev_tx_async_split_legacy(dev, vq, mbuf_pool, pkts,
+ count, dma_id, vchan_id);
+ else
+ count = virtio_dev_tx_async_split_compliant(dev, vq, mbuf_pool, pkts,
+ count, dma_id, vchan_id);
+
+ *nr_inflight = vq->async->pkts_inflight_n;
+
+out:
+ if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
+ vhost_user_iotlb_rd_unlock(vq);
+
+out_access_unlock:
+ rte_spinlock_unlock(&vq->access_lock);
+
+ if (unlikely(rarp_mbuf != NULL))
+ count += 1;
+
+ return count;
+}