Skip to contentSkip to conversation

vhost: fix async packed ring batch datapath

PATCH Applied Cheng Jiang 3 participants 3 messages Jul 8, 2021 – Jul 20, 2021

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 vhost: fix async packed ring batch datapath

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")
---
 lib/vhost/virtio_net.c | 163 ++++++++++++++++++++++++++++++++---------
 1 file changed, 128 insertions(+), 35 deletions(-)
lib/vhost/virtio_net.c +128 -35
index 6bd00b746b..f4a2c88d8b 100644
@@ -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,
Signature · 1 line
2.29.2
Reviewed-by on v1 1/1
#1 lore raw

Hi,

Quoted 9 lines · Cheng Jiang

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.

Quoted 8 lines
Fixes 873e8dad6f49 ("vhost: support packed ring in async datapath")
---
 lib/vhost/virtio_net.c | 163 ++++++++++++++++++++++++++++++++---------
 1 file changed, 128 insertions(+), 35 deletions(-)

Thanks,
Maxime

Applied
#2 lore raw
In reply to Cheng Jiang
Quoted 18 lines

-----Original Message-----
From: Jiang, Cheng1 <[email protected]>
Sent: Thursday, July 8, 2021 6:25 PM
To: [email protected]; Xia, Chenbo <[email protected]>
Cc: [email protected]; Hu, Jiayu <[email protected]>; Yang, YvonneX <[email protected]>; Jiang, Cheng1 <[email protected]>; [email protected]
Subject: [PATCH] vhost: fix async packed ring batch datapath

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.

Quoted 7 lines
Fixes 873e8dad6f49 ("vhost: support packed ring in async datapath")

---
--
2.29.2

Applied to next-virtio/main, thanks.