virtio-net: support XDP when not more queues

The number of queues implemented by many virtio backends is limited,
especially some machines have a large number of CPUs. In this case, it
is often impossible to allocate a separate queue for
XDP_TX/XDP_REDIRECT, then xdp cannot be loaded to work, even xdp does
not use the XDP_TX/XDP_REDIRECT.

This patch allows XDP_TX/XDP_REDIRECT to run by reuse the existing SQ
with __netif_tx_lock() hold when there are not enough queues.
Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Reviewed-by: Dust Li <dust.li@linux.alibaba.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/virtio_net.c

@@ -195,6 +195,9 @@ struct virtnet_info {
 	/* # of XDP queue pairs currently used by the driver */
 	u16 xdp_queue_pairs;
 
+	/* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
+	bool xdp_enabled;
+
 	/* I like... big packets and I cannot lie! */
 	bool big_packets;
 
@@ -481,12 +484,41 @@ static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
 	return 0;
 }
 
-static struct send_queue *virtnet_xdp_sq(struct virtnet_info *vi)
-{
-	unsigned int qp;
-
-	qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
-	return &vi->sq[qp];
+/* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
+ * the current cpu, so it does not need to be locked.
+ *
+ * Here we use marco instead of inline functions because we have to deal with
+ * three issues at the same time: 1. the choice of sq. 2. judge and execute the
+ * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
+ * functions to perfectly solve these three problems at the same time.
+ */
+#define virtnet_xdp_get_sq(vi) ({                                       \
+	struct netdev_queue *txq;                                       \
+	typeof(vi) v = (vi);                                            \
+	unsigned int qp;                                                \
+									\
+	if (v->curr_queue_pairs > nr_cpu_ids) {                         \
+		qp = v->curr_queue_pairs - v->xdp_queue_pairs;          \
+		qp += smp_processor_id();                               \
+		txq = netdev_get_tx_queue(v->dev, qp);                  \
+		__netif_tx_acquire(txq);                                \
+	} else {                                                        \
+		qp = smp_processor_id() % v->curr_queue_pairs;          \
+		txq = netdev_get_tx_queue(v->dev, qp);                  \
+		__netif_tx_lock(txq, raw_smp_processor_id());           \
+	}                                                               \
+	v->sq + qp;                                                     \
+})
+
+#define virtnet_xdp_put_sq(vi, q) {                                     \
+	struct netdev_queue *txq;                                       \
+	typeof(vi) v = (vi);                                            \
+									\
+	txq = netdev_get_tx_queue(v->dev, (q) - v->sq);                 \
+	if (v->curr_queue_pairs > nr_cpu_ids)                           \
+		__netif_tx_release(txq);                                \
+	else                                                            \
+		__netif_tx_unlock(txq);                                 \
 }
 
 static int virtnet_xdp_xmit(struct net_device *dev,
@@ -512,7 +544,7 @@ static int virtnet_xdp_xmit(struct net_device *dev,
 	if (!xdp_prog)
 		return -ENXIO;
 
-	sq = virtnet_xdp_sq(vi);
+	sq = virtnet_xdp_get_sq(vi);
 
 	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
 		ret = -EINVAL;
@@ -560,12 +592,13 @@ static int virtnet_xdp_xmit(struct net_device *dev,
 	sq->stats.kicks += kicks;
 	u64_stats_update_end(&sq->stats.syncp);
 
+	virtnet_xdp_put_sq(vi, sq);
 	return ret;
 }
 
 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 {
-	return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
+	return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
 }
 
 /* We copy the packet for XDP in the following cases:
@@ -1458,12 +1491,13 @@ static int virtnet_poll(struct napi_struct *napi, int budget)
 		xdp_do_flush();
 
 	if (xdp_xmit & VIRTIO_XDP_TX) {
-		sq = virtnet_xdp_sq(vi);
+		sq = virtnet_xdp_get_sq(vi);
 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
 			u64_stats_update_begin(&sq->stats.syncp);
 			sq->stats.kicks++;
 			u64_stats_update_end(&sq->stats.syncp);
 		}
+		virtnet_xdp_put_sq(vi, sq);
 	}
 
 	return received;
@@ -2418,10 +2452,9 @@ static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
 
 	/* XDP requires extra queues for XDP_TX */
 	if (curr_qp + xdp_qp > vi->max_queue_pairs) {
-		NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
-		netdev_warn(dev, "request %i queues but max is %i\n",
+		netdev_warn(dev, "XDP request %i queues but max is %i. XDP_TX and XDP_REDIRECT will operate in a slower locked tx mode.\n",
 			    curr_qp + xdp_qp, vi->max_queue_pairs);
-		return -ENOMEM;
+		xdp_qp = 0;
 	}
 
 	old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
@@ -2455,11 +2488,14 @@ static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
 	vi->xdp_queue_pairs = xdp_qp;
 
 	if (prog) {
+		vi->xdp_enabled = true;
 		for (i = 0; i < vi->max_queue_pairs; i++) {
 			rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
 			if (i == 0 && !old_prog)
 				virtnet_clear_guest_offloads(vi);
 		}
+	} else {
+		vi->xdp_enabled = false;
 	}
 
 	for (i = 0; i < vi->max_queue_pairs; i++) {
@@ -2527,7 +2563,7 @@ static int virtnet_set_features(struct net_device *dev,
 	int err;
 
 	if ((dev->features ^ features) & NETIF_F_LRO) {
-		if (vi->xdp_queue_pairs)
+		if (vi->xdp_enabled)
 			return -EBUSY;
 
 		if (features & NETIF_F_LRO)