[PATCH net-next v2] netdevsim: implement peer queue flow control

[PATCH net-next v2] netdevsim: implement peer queue flow control

[Breno Leitao]

Add flow control mechanism between paired netdevsim devices to stop the
TX queue during high traffic scenarios. When a receive queue becomes
congested (approaching NSIM_RING_SIZE limit), the corresponding transmit
queue on the peer device is stopped using netif_subqueue_try_stop().

Once the receive queue has sufficient capacity again, the peer's
transmit queue is resumed with netif_tx_wake_queue().

Key changes:
  * Add nsim_stop_peer_tx_queue() to pause peer TX when RX queue is full
  * Add nsim_start_peer_tx_queue() to resume peer TX when RX queue drains
  * Implement queue mapping validation to ensure TX/RX queue count match
  * Wake all queues during device unlinking to prevent stuck queues
  * Use RCU protection when accessing peer device references

The flow control only activates when devices have matching TX/RX queue
counts to ensure proper queue mapping.

Suggested-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Breno Leitao <leitao@debian.org>
---
Changes in v2:
- Move the RCU locks inside the function (David)
- Use better helpers for waking up all the queues (Jakub)
- Link to v1: https://lore.kernel.org/r/20250701-netdev_flow_control-v1-1-240329fc91b1@debian.org
---
 drivers/net/netdevsim/bus.c    |  3 ++
 drivers/net/netdevsim/netdev.c | 63 ++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 64 insertions(+), 2 deletions(-)

diff --git a/drivers/net/netdevsim/bus.c b/drivers/net/netdevsim/bus.c
index 64c0cdd31bf85..1ba52471f3fbc 100644
--- a/drivers/net/netdevsim/bus.c
+++ b/drivers/net/netdevsim/bus.c
@@ -366,6 +366,9 @@ static ssize_t unlink_device_store(const struct bus_type *bus, const char *buf,
 	err = 0;
 	RCU_INIT_POINTER(nsim->peer, NULL);
 	RCU_INIT_POINTER(peer->peer, NULL);
+	synchronize_net();
+	netif_tx_wake_all_queues(dev);
+	netif_tx_wake_all_queues(peer->netdev);
 
 out_put_netns:
 	put_net(ns);
diff --git a/drivers/net/netdevsim/netdev.c b/drivers/net/netdevsim/netdev.c
index e36d3e846c2dc..b5b13fba6450d 100644
--- a/drivers/net/netdevsim/netdev.c
+++ b/drivers/net/netdevsim/netdev.c
@@ -37,9 +37,67 @@ MODULE_IMPORT_NS("NETDEV_INTERNAL");
 
 #define NSIM_RING_SIZE		256
 
-static int nsim_napi_rx(struct nsim_rq *rq, struct sk_buff *skb)
+static void nsim_start_peer_tx_queue(struct net_device *dev, struct nsim_rq *rq)
+{
+	struct netdevsim *ns = netdev_priv(dev);
+	struct net_device *peer_dev;
+	struct netdevsim *peer_ns;
+	struct netdev_queue *txq;
+	u16 idx;
+
+	idx = rq->napi.index;
+	rcu_read_lock();
+	peer_ns = rcu_dereference(ns->peer);
+	if (!peer_ns)
+		goto out;
+
+	/* TX device */
+	peer_dev = peer_ns->netdev;
+	if (dev->real_num_tx_queues != peer_dev->num_rx_queues)
+		goto out;
+
+	txq = netdev_get_tx_queue(peer_dev, idx);
+	if (!(netif_tx_queue_stopped(txq)))
+		goto out;
+
+	netif_tx_wake_queue(txq);
+out:
+	rcu_read_unlock();
+}
+
+static void nsim_stop_peer_tx_queue(struct net_device *dev, struct nsim_rq *rq,
+				    u16 idx)
+{
+	struct netdevsim *ns = netdev_priv(dev);
+	struct net_device *peer_dev;
+	struct netdevsim *peer_ns;
+
+	rcu_read_lock();
+	peer_ns = rcu_dereference(ns->peer);
+	if (!peer_ns)
+		goto out;
+
+	/* TX device */
+	peer_dev = peer_ns->netdev;
+
+	/* If different queues size, do not stop, since it is not
+	 * easy to find which TX queue is mapped here
+	 */
+	if (dev->real_num_tx_queues != peer_dev->num_rx_queues)
+		goto out;
+
+	netif_subqueue_try_stop(peer_dev, idx,
+				NSIM_RING_SIZE - skb_queue_len(&rq->skb_queue),
+				NSIM_RING_SIZE / 2);
+out:
+	rcu_read_unlock();
+}
+
+static int nsim_napi_rx(struct net_device *dev, struct nsim_rq *rq,
+			struct sk_buff *skb)
 {
 	if (skb_queue_len(&rq->skb_queue) > NSIM_RING_SIZE) {
+		nsim_stop_peer_tx_queue(dev, rq, skb_get_queue_mapping(skb));
 		dev_kfree_skb_any(skb);
 		return NET_RX_DROP;
 	}
@@ -51,7 +109,7 @@ static int nsim_napi_rx(struct nsim_rq *rq, struct sk_buff *skb)
 static int nsim_forward_skb(struct net_device *dev, struct sk_buff *skb,
 			    struct nsim_rq *rq)
 {
-	return __dev_forward_skb(dev, skb) ?: nsim_napi_rx(rq, skb);
+	return __dev_forward_skb(dev, skb) ?: nsim_napi_rx(dev, rq, skb);
 }
 
 static netdev_tx_t nsim_start_xmit(struct sk_buff *skb, struct net_device *dev)
@@ -351,6 +409,7 @@ static int nsim_rcv(struct nsim_rq *rq, int budget)
 			dev_dstats_rx_dropped(dev);
 	}
 
+	nsim_start_peer_tx_queue(dev, rq);
 	return i;
 }
 

---
base-commit: be4ea6c336b9a4fc1cc4be1c0549b24d0e687488
change-id: 20250630-netdev_flow_control-2b2d37965377

Best regards,
--  
Breno Leitao <leitao@debian.org>

Re: [PATCH net-next v2] netdevsim: implement peer queue flow control

[Jakub Kicinski]

On Thu, 03 Jul 2025 06:09:31 -0700 Breno Leitao wrote:
> +static int nsim_napi_rx(struct net_device *dev, struct nsim_rq *rq,
> +			struct sk_buff *skb)
>  {
>  	if (skb_queue_len(&rq->skb_queue) > NSIM_RING_SIZE) {
> +		nsim_stop_peer_tx_queue(dev, rq, skb_get_queue_mapping(skb));
>  		dev_kfree_skb_any(skb);
>  		return NET_RX_DROP;
>  	}

we should probably add:

	if (skb_queue_len(&rq->skb_queue) > NSIM_RING_SIZE)
		nsim_stop_tx_queue(dev, rq, skb_get_queue_mapping(skb));

after enqueuing the skb, so that we stop the queue before any drops
happen

> @@ -51,7 +109,7 @@ static int nsim_napi_rx(struct nsim_rq *rq, struct sk_buff *skb)
>  static int nsim_forward_skb(struct net_device *dev, struct sk_buff *skb,
>  			    struct nsim_rq *rq)
>  {
> -	return __dev_forward_skb(dev, skb) ?: nsim_napi_rx(rq, skb);
> +	return __dev_forward_skb(dev, skb) ?: nsim_napi_rx(dev, rq, skb);
>  }
>  
>  static netdev_tx_t nsim_start_xmit(struct sk_buff *skb, struct net_device *dev)

nsim_start_xmit() has both dev and peer_dev, pass them all the way to
nsim_stop_peer_tx_queue() so that you don't have to try to dereference
the peer again.

> +	if (dev->real_num_tx_queues != peer_dev->num_rx_queues)

given that we compare real_num_tx_queues I think we should also kick
the queues in nsim_set_channels(), like we do in unlink_device_store()
-- 
pw-bot: cr

Re: [PATCH net-next v2] netdevsim: implement peer queue flow control

[Breno Leitao]

Hello Jakub,

On Tue, Jul 08, 2025 at 06:27:18PM -0700, Jakub Kicinski wrote:
> On Thu, 03 Jul 2025 06:09:31 -0700 Breno Leitao wrote:
> > +static int nsim_napi_rx(struct net_device *dev, struct nsim_rq *rq,
> > +			struct sk_buff *skb)
> >  {
> >  	if (skb_queue_len(&rq->skb_queue) > NSIM_RING_SIZE) {
> > +		nsim_stop_peer_tx_queue(dev, rq, skb_get_queue_mapping(skb));
> >  		dev_kfree_skb_any(skb);
> >  		return NET_RX_DROP;
> >  	}
> 
> we should probably add:
> 
> 	if (skb_queue_len(&rq->skb_queue) > NSIM_RING_SIZE)
> 		nsim_stop_tx_queue(dev, rq, skb_get_queue_mapping(skb));
> 
> after enqueuing the skb, so that we stop the queue before any drops
> happen

Agree, we can stop the queue when queueing the packets instead. Since we
need to check for the queue numbers, we cannot call nsim_stop_tx_queue()
straight away. I think we still need to have a helper
(nsim_stop_tx_queue). This is what I have in mind:

	static void nsim_stop_tx_queue(struct net_device *tx_dev,
					struct net_device *rx_dev,
					struct nsim_rq *rq,
					u16 idx)
	{
		/* If different queues size, do not stop, since it is not
		* easy to find which TX queue is mapped here
		*/
		if (rx_dev->real_num_tx_queues != tx_dev->num_rx_queues)
			return;

		/* rq is the queue on the receive side */
		netif_subqueue_try_stop(tx_dev, idx,
					NSIM_RING_SIZE - skb_queue_len(&rq->skb_queue),
					NSIM_RING_SIZE / 2);
	}

	static int nsim_napi_rx(struct net_device *tx_dev, struct net_device *rx_dev,
				struct nsim_rq *rq, struct sk_buff *skb)
	{
		if (skb_queue_len(&rq->skb_queue) > NSIM_RING_SIZE) {
			dev_kfree_skb_any(skb);
			return NET_RX_DROP;
		}

		skb_queue_tail(&rq->skb_queue, skb);

		/* Stop the peer TX queue avoiding dropping packets later */
		if (skb_queue_len(&rq->skb_queue) >= NSIM_RING_SIZE)
			nsim_stop_tx_queue(tx_dev, rx_dev, rq,
					skb_get_queue_mapping(skb));

		return NET_RX_SUCCESS;
	}

> > @@ -51,7 +109,7 @@ static int nsim_napi_rx(struct nsim_rq *rq, struct sk_buff *skb)
> >  static int nsim_forward_skb(struct net_device *dev, struct sk_buff *skb,
> >  			    struct nsim_rq *rq)
> >  {
> > -	return __dev_forward_skb(dev, skb) ?: nsim_napi_rx(rq, skb);
> > +	return __dev_forward_skb(dev, skb) ?: nsim_napi_rx(dev, rq, skb);
> >  }
> >  
> >  static netdev_tx_t nsim_start_xmit(struct sk_buff *skb, struct net_device *dev)
> 
> nsim_start_xmit() has both dev and peer_dev, pass them all the way to
> nsim_stop_peer_tx_queue() so that you don't have to try to dereference
> the peer again.

Sure. This is a good idea. I am using it, as you can see in the snippet
above.

> > +	if (dev->real_num_tx_queues != peer_dev->num_rx_queues)
> 
> given that we compare real_num_tx_queues I think we should also kick
> the queues in nsim_set_channels(), like we do in unlink_device_store()

Sure. I suppose something like the following. What do you think?

	nsim_set_channels(struct net_device *dev, struct ethtool_channels *ch)
	{
		struct netdevsim *ns = netdev_priv(dev);
	+       struct netdevsim *peer;
		int err;

		err = netif_set_real_num_queues(dev, ch->combined_count,
	@@ -113,6 +114,14 @@ nsim_set_channels(struct net_device *dev, struct ethtool_channels *ch)
			return err;

		ns->ethtool.channels = ch->combined_count;
	+
	+	synchronize_net();
	+       netif_tx_wake_all_queues(dev);
	+       rcu_read_lock();
	+       peer = rcu_dereference(ns->peer);
	+       if (peer)
	+               netif_tx_wake_all_queues(peer->netdev);
	+       rcu_read_unlock();
	+
		return 0;
	}


Also, with this patch, we will eventually get the following critical
message:

	net_crit_ratelimited("Virtual device %s asks to queue packet!\n", dev->name);

I am wondering if that alert is not valid anymore, and I can simply
remove it.

Thanks for your review!
--breno

Re: [PATCH net-next v2] netdevsim: implement peer queue flow control

[Jakub Kicinski]

On Wed, 9 Jul 2025 03:34:20 -0700 Breno Leitao wrote:

> On Tue, Jul 08, 2025 at 06:27:18PM -0700, Jakub Kicinski wrote:
> > On Thu, 03 Jul 2025 06:09:31 -0700 Breno Leitao wrote:  
> > > +static int nsim_napi_rx(struct net_device *dev, struct nsim_rq *rq,
> > > +			struct sk_buff *skb)
> > >  {
> > >  	if (skb_queue_len(&rq->skb_queue) > NSIM_RING_SIZE) {
> > > +		nsim_stop_peer_tx_queue(dev, rq, skb_get_queue_mapping(skb));
> > >  		dev_kfree_skb_any(skb);
> > >  		return NET_RX_DROP;
> > >  	}  
> > 
> > we should probably add:
> > 
> > 	if (skb_queue_len(&rq->skb_queue) > NSIM_RING_SIZE)
> > 		nsim_stop_tx_queue(dev, rq, skb_get_queue_mapping(skb));
> > 
> > after enqueuing the skb, so that we stop the queue before any drops
> > happen  
> 
> Agree, we can stop the queue when queueing the packets instead. Since we
> need to check for the queue numbers, we cannot call nsim_stop_tx_queue()
> straight away. I think we still need to have a helper
> (nsim_stop_tx_queue). This is what I have in mind:

LGTM!

> > > +	if (dev->real_num_tx_queues != peer_dev->num_rx_queues)  
> > 
> > given that we compare real_num_tx_queues I think we should also kick
> > the queues in nsim_set_channels(), like we do in unlink_device_store()  
> 
> Sure. I suppose something like the following. What do you think?
> 
> 	nsim_set_channels(struct net_device *dev, struct ethtool_channels *ch)
> 	{
> 		struct netdevsim *ns = netdev_priv(dev);
> 	+       struct netdevsim *peer;
> 		int err;
> 
> 		err = netif_set_real_num_queues(dev, ch->combined_count,
> 	@@ -113,6 +114,14 @@ nsim_set_channels(struct net_device *dev, struct ethtool_channels *ch)
> 			return err;
> 
> 		ns->ethtool.channels = ch->combined_count;
> 	+
> 	+	synchronize_net();
> 	+       netif_tx_wake_all_queues(dev);
> 	+       rcu_read_lock();
> 	+       peer = rcu_dereference(ns->peer);
> 	+       if (peer)
> 	+               netif_tx_wake_all_queues(peer->netdev);
> 	+       rcu_read_unlock();

That's sufficiently orthogonal to warrant a dedicated function / helper.

In terms of code I think we can skip the whole dance if peer is NULL?

> 		return 0;
> 	}
> 
> 
> Also, with this patch, we will eventually get the following critical
> message:
> 
> 	net_crit_ratelimited("Virtual device %s asks to queue packet!\n", dev->name);
> 
> I am wondering if that alert is not valid anymore, and I can simply
> remove it.

Ah. In nsim_setup() we should remove IFF_NO_QUEUE and stop setting
tx_queue_len to 0

Re: [PATCH net-next v2] netdevsim: implement peer queue flow control

[Breno Leitao]

On Wed, Jul 09, 2025 at 02:36:27PM -0700, Jakub Kicinski wrote:
> On Wed, 9 Jul 2025 03:34:20 -0700 Breno Leitao wrote:
> > 	+
> > 	+	synchronize_net();
> > 	+       netif_tx_wake_all_queues(dev);
> > 	+       rcu_read_lock();
> > 	+       peer = rcu_dereference(ns->peer);
> > 	+       if (peer)
> > 	+               netif_tx_wake_all_queues(peer->netdev);
> > 	+       rcu_read_unlock();
> 
> That's sufficiently orthogonal to warrant a dedicated function / helper.
> 
> In terms of code I think we can skip the whole dance if peer is NULL?

Sure. We can use rcu_access_pointer() to check if the value is set, and
then get into the slow path.

       if (rcu_access_pointer(ns->peer))
               nsim_wake_queues(dev);

> > Also, with this patch, we will eventually get the following critical
> > message:
> > 
> > 	net_crit_ratelimited("Virtual device %s asks to queue packet!\n", dev->name);
> > 
> > I am wondering if that alert is not valid anymore, and I can simply
> > remove it.
> 
> Ah. In nsim_setup() we should remove IFF_NO_QUEUE and stop setting
> tx_queue_len to 0

That makes sense, thanks!
--breno