vhost: fix async packed ring batch datapath
authorCheng Jiang <cheng1.jiang@intel.com>
Thu, 8 Jul 2021 10:25:28 +0000 (10:25 +0000)
committerChenbo Xia <chenbo.xia@intel.com>
Wed, 21 Jul 2021 05:56:13 +0000 (07:56 +0200)
We assume that in the sync path, if there is no buffer wrap in the
avail descriptors fetched in a batch, there is no buffer wrap in the
used descriptors which need to be written back in this batch, but
this assumption is wrong in the async path since there are inflight
descriptors which are processed by the DMA device.

This patch refactors the batch copy code and adds used ring buffer
wrap check as a batch copy condition to fix this issue.

Fixes: 873e8dad6f49 ("vhost: support packed ring in async datapath")
Cc: stable@dpdk.org
Signed-off-by: Cheng Jiang <cheng1.jiang@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
lib/vhost/virtio_net.c

index 7dba334..f0b78f5 100644 (file)
@@ -221,11 +221,6 @@ vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
        uint16_t last_used_idx;
        struct vring_packed_desc *desc_base;
 
-       if (vq->shadow_used_idx) {
-               do_data_copy_enqueue(dev, vq);
-               vhost_flush_enqueue_shadow_packed(dev, vq);
-       }
-
        last_used_idx = vq->last_used_idx;
        desc_base = &vq->desc_packed[last_used_idx];
 
@@ -1258,18 +1253,16 @@ virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 }
 
 static __rte_always_inline int
-virtio_dev_rx_batch_packed(struct virtio_net *dev,
+virtio_dev_rx_sync_batch_check(struct virtio_net *dev,
                           struct vhost_virtqueue *vq,
-                          struct rte_mbuf **pkts)
+                          struct rte_mbuf **pkts,
+                          uint64_t *desc_addrs,
+                          uint64_t *lens)
 {
        bool wrap_counter = vq->avail_wrap_counter;
        struct vring_packed_desc *descs = vq->desc_packed;
        uint16_t avail_idx = vq->last_avail_idx;
-       uint64_t desc_addrs[PACKED_BATCH_SIZE];
-       struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
        uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
-       uint64_t lens[PACKED_BATCH_SIZE];
-       uint16_t ids[PACKED_BATCH_SIZE];
        uint16_t i;
 
        if (unlikely(avail_idx & PACKED_BATCH_MASK))
@@ -1307,6 +1300,84 @@ virtio_dev_rx_batch_packed(struct virtio_net *dev,
                        return -1;
        }
 
+       return 0;
+}
+
+static __rte_always_inline int
+virtio_dev_rx_async_batch_check(struct virtio_net *dev,
+                          struct vhost_virtqueue *vq,
+                          struct rte_mbuf **pkts,
+                          uint64_t *desc_addrs,
+                          uint64_t *lens)
+{
+       bool wrap_counter = vq->avail_wrap_counter;
+       struct vring_packed_desc *descs = vq->desc_packed;
+       uint16_t avail_idx = vq->last_avail_idx;
+       uint16_t used_idx = vq->last_used_idx;
+       uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
+       uint32_t cpy_threshold = vq->async_threshold;
+       uint16_t i;
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               if (unlikely(pkts[i]->data_len >= cpy_threshold))
+                       return -1;
+       }
+
+       if (unlikely(avail_idx & PACKED_BATCH_MASK))
+               return -1;
+
+       if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
+               return -1;
+
+       if (unlikely((used_idx + PACKED_BATCH_SIZE) > vq->size))
+               return -1;
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               if (unlikely(pkts[i]->next != NULL))
+                       return -1;
+               if (unlikely(!desc_is_avail(&descs[avail_idx + i],
+                                           wrap_counter)))
+                       return -1;
+       }
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+               lens[i] = descs[avail_idx + i].len;
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
+                       return -1;
+       }
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+               desc_addrs[i] = vhost_iova_to_vva(dev, vq,
+                                                 descs[avail_idx + i].addr,
+                                                 &lens[i],
+                                                 VHOST_ACCESS_RW);
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               if (unlikely(!desc_addrs[i]))
+                       return -1;
+               if (unlikely(lens[i] != descs[avail_idx + i].len))
+                       return -1;
+       }
+
+       return 0;
+}
+
+static __rte_always_inline void
+virtio_dev_rx_batch_packed_copy(struct virtio_net *dev,
+                          struct vhost_virtqueue *vq,
+                          struct rte_mbuf **pkts,
+                          uint64_t *desc_addrs,
+                          uint64_t *lens)
+{
+       uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
+       struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
+       struct vring_packed_desc *descs = vq->desc_packed;
+       uint16_t avail_idx = vq->last_avail_idx;
+       uint16_t ids[PACKED_BATCH_SIZE];
+       uint16_t i;
+
        vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
                rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
                hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
@@ -1334,6 +1405,51 @@ virtio_dev_rx_batch_packed(struct virtio_net *dev,
                ids[i] = descs[avail_idx + i].id;
 
        vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
+}
+
+static __rte_always_inline int
+virtio_dev_rx_sync_batch_packed(struct virtio_net *dev,
+                          struct vhost_virtqueue *vq,
+                          struct rte_mbuf **pkts)
+{
+       uint64_t desc_addrs[PACKED_BATCH_SIZE];
+       uint64_t lens[PACKED_BATCH_SIZE];
+
+       if (virtio_dev_rx_sync_batch_check(dev, vq, pkts, desc_addrs, lens) == -1)
+               return -1;
+
+       if (vq->shadow_used_idx) {
+               do_data_copy_enqueue(dev, vq);
+               vhost_flush_enqueue_shadow_packed(dev, vq);
+       }
+
+       virtio_dev_rx_batch_packed_copy(dev, vq, pkts, desc_addrs, lens);
+
+       return 0;
+}
+
+static __rte_always_inline int
+virtio_dev_rx_async_batch_packed(struct virtio_net *dev,
+                          struct vhost_virtqueue *vq,
+                          struct rte_mbuf **pkts,
+                          struct rte_mbuf **comp_pkts, uint32_t *pkt_done)
+{
+       uint16_t i;
+       uint64_t desc_addrs[PACKED_BATCH_SIZE];
+       uint64_t lens[PACKED_BATCH_SIZE];
+
+       if (virtio_dev_rx_async_batch_check(dev, vq, pkts, desc_addrs, lens) == -1)
+               return -1;
+
+       virtio_dev_rx_batch_packed_copy(dev, vq, pkts, desc_addrs, lens);
+
+       if (vq->shadow_used_idx) {
+               do_data_copy_enqueue(dev, vq);
+               vhost_flush_enqueue_shadow_packed(dev, vq);
+       }
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+               comp_pkts[(*pkt_done)++] = pkts[i];
 
        return 0;
 }
@@ -1375,7 +1491,7 @@ virtio_dev_rx_packed(struct virtio_net *dev,
                rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
 
                if (count - pkt_idx >= PACKED_BATCH_SIZE) {
-                       if (!virtio_dev_rx_batch_packed(dev, vq,
+                       if (!virtio_dev_rx_sync_batch_packed(dev, vq,
                                                        &pkts[pkt_idx])) {
                                pkt_idx += PACKED_BATCH_SIZE;
                                continue;
@@ -1728,29 +1844,6 @@ vhost_update_used_packed(struct vhost_virtqueue *vq,
        vq->desc_packed[head_idx].flags = head_flags;
 }
 
-static __rte_always_inline int
-virtio_dev_rx_async_batch_packed(struct virtio_net *dev,
-                          struct vhost_virtqueue *vq,
-                          struct rte_mbuf **pkts,
-                          struct rte_mbuf **comp_pkts, uint32_t *pkt_done)
-{
-       uint16_t i;
-       uint32_t cpy_threshold = vq->async_threshold;
-
-       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
-               if (unlikely(pkts[i]->pkt_len >= cpy_threshold))
-                       return -1;
-       }
-       if (!virtio_dev_rx_batch_packed(dev, vq, pkts)) {
-               vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
-                       comp_pkts[(*pkt_done)++] = pkts[i];
-
-               return 0;
-       }
-
-       return -1;
-}
-
 static __rte_always_inline int
 vhost_enqueue_async_single_packed(struct virtio_net *dev,
                            struct vhost_virtqueue *vq,