vhost: simplify async enqueue completion
authorMaxime Coquelin <maxime.coquelin@redhat.com>
Tue, 26 Oct 2021 16:28:59 +0000 (18:28 +0200)
committerMaxime Coquelin <maxime.coquelin@redhat.com>
Fri, 29 Oct 2021 10:32:30 +0000 (12:32 +0200)
vhost_poll_enqueue_completed() assumes some inflight
packets could have been completed in a previous call but
not returned to the application. But this is not the case,
since check_completed_copies callback is never called with
more than the current count as argument.

In other words, async->last_pkts_n is always 0. Removing it
greatly simplifies the function.

Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Reviewed-by: Jiayu Hu <jiayu.hu@intel.com>
lib/vhost/vhost.h
lib/vhost/virtio_net.c

index 1e3f1ba..1ba9a60 100644 (file)
@@ -141,7 +141,6 @@ struct vhost_async {
        struct async_inflight_info *pkts_info;
        uint16_t pkts_idx;
        uint16_t pkts_inflight_n;
-       uint16_t last_pkts_n;
        union {
                struct vring_used_elem  *descs_split;
                struct vring_used_elem_packed *buffers_packed;
index dfe864b..3206b3f 100644 (file)
@@ -1619,7 +1619,11 @@ virtio_dev_rx_async_submit_split(struct virtio_net *dev,
                                vq->shadow_used_idx);
 
                async->desc_idx_split += vq->shadow_used_idx;
+
                async->pkts_idx += pkt_idx;
+               if (async->pkts_idx >= vq->size)
+                       async->pkts_idx -= vq->size;
+
                async->pkts_inflight_n += pkt_idx;
                vq->shadow_used_idx = 0;
        }
@@ -1921,68 +1925,44 @@ static __rte_always_inline uint16_t
 vhost_poll_enqueue_completed(struct virtio_net *dev, uint16_t queue_id,
                struct rte_mbuf **pkts, uint16_t count)
 {
-       struct vhost_virtqueue *vq;
-       struct vhost_async *async;
-       struct async_inflight_info *pkts_info;
+       struct vhost_virtqueue *vq = dev->virtqueue[queue_id];
+       struct vhost_async *async = vq->async;
+       struct async_inflight_info *pkts_info = async->pkts_info;
        int32_t n_cpl;
-       uint16_t n_pkts_cpl = 0, n_pkts_put = 0, n_descs = 0, n_buffers = 0;
-       uint16_t start_idx, pkts_idx, vq_size;
-       uint16_t from, i;
+       uint16_t n_descs = 0, n_buffers = 0;
+       uint16_t start_idx, from, i;
 
-       vq = dev->virtqueue[queue_id];
-       async = vq->async;
-       pkts_idx = async->pkts_idx % vq->size;
-       pkts_info = async->pkts_info;
-       vq_size = vq->size;
-       start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
-               vq_size, async->pkts_inflight_n);
-
-       if (count > async->last_pkts_n) {
-               n_cpl = async->ops.check_completed_copies(dev->vid,
-                       queue_id, 0, count - async->last_pkts_n);
-               if (likely(n_cpl >= 0)) {
-                       n_pkts_cpl = n_cpl;
-               } else {
-                       VHOST_LOG_DATA(ERR,
-                               "(%d) %s: failed to check completed copies for queue id %d.\n",
+       start_idx = virtio_dev_rx_async_get_info_idx(async->pkts_idx,
+               vq->size, async->pkts_inflight_n);
+
+       n_cpl = async->ops.check_completed_copies(dev->vid, queue_id, 0, count);
+       if (unlikely(n_cpl < 0)) {
+               VHOST_LOG_DATA(ERR, "(%d) %s: failed to check completed copies for queue id %d.\n",
                                dev->vid, __func__, queue_id);
-                       n_pkts_cpl = 0;
-               }
+               return 0;
        }
 
-       n_pkts_cpl += async->last_pkts_n;
-       n_pkts_put = RTE_MIN(n_pkts_cpl, count);
-       if (unlikely(n_pkts_put == 0)) {
-               async->last_pkts_n = n_pkts_cpl;
+       if (n_cpl == 0)
                return 0;
-       }
 
-       if (vq_is_packed(dev)) {
-               for (i = 0; i < n_pkts_put; i++) {
-                       from = (start_idx + i) % vq_size;
-                       n_buffers += pkts_info[from].nr_buffers;
-                       pkts[i] = pkts_info[from].mbuf;
-               }
-       } else {
-               for (i = 0; i < n_pkts_put; i++) {
-                       from = (start_idx + i) & (vq_size - 1);
-                       n_descs += pkts_info[from].descs;
-                       pkts[i] = pkts_info[from].mbuf;
-               }
+       for (i = 0; i < n_cpl; i++) {
+               from = (start_idx + i) % vq->size;
+               /* Only used with packed ring */
+               n_buffers += pkts_info[from].nr_buffers;
+               /* Only used with split ring */
+               n_descs += pkts_info[from].descs;
+               pkts[i] = pkts_info[from].mbuf;
        }
-       async->last_pkts_n = n_pkts_cpl - n_pkts_put;
-       async->pkts_inflight_n -= n_pkts_put;
+
+       async->pkts_inflight_n -= n_cpl;
 
        if (likely(vq->enabled && vq->access_ok)) {
                if (vq_is_packed(dev)) {
                        write_back_completed_descs_packed(vq, n_buffers);
-
                        vhost_vring_call_packed(dev, vq);
                } else {
                        write_back_completed_descs_split(vq, n_descs);
-
-                       __atomic_add_fetch(&vq->used->idx, n_descs,
-                                       __ATOMIC_RELEASE);
+                       __atomic_add_fetch(&vq->used->idx, n_descs, __ATOMIC_RELEASE);
                        vhost_vring_call_split(dev, vq);
                }
        } else {
@@ -1995,7 +1975,7 @@ vhost_poll_enqueue_completed(struct virtio_net *dev, uint16_t queue_id,
                }
        }
 
-       return n_pkts_put;
+       return n_cpl;
 }
 
 uint16_t