veth: allow GRO even without XDP

After the previous patch, when enabling GRO, locally generated TCP traffic experiences some measurable overhead, as it traverses the GRO engine without any chance of aggregation.

This change refine the NAPI receive path admission test, to avoid unnecessary GRO overhead in most scenarios, when GRO is enabled on a veth peer.

Only skbs that are eligible for aggregation enter the GRO layer, the others will go through the traditional receive path.

Signed-off-by Paolo Abeni <[email protected]>

---
 drivers/net/veth.c | 23 ++++++++++++++++++++++-
 1 file changed, 22 insertions(+), 1 deletion(-)

drivers/net/veth.c +19 -0

index ca44e82d1edeb..85f90f33d437e 100644
@@ -282,6 +282,25 @@ static int veth_forward_skb(struct net_device *dev, struct sk_buff *skb,
 		netif_rx(skb);
 }
 
+/* return true if the specified skb has chances of GRO aggregation
+ * Don't strive for accuracy, but try to avoid GRO overhead in the most
+ * common scenarios.
+ * When XDP is enabled, all traffic is considered eligible, as the xmit
+ * device has TSO off.
+ * When TSO is enabled on the xmit device, we are likely interested only
+ * in UDP aggregation, explicitly check for that if the skb is suspected
+ * - the sock_wfree destructor is used by UDP, ICMP and XDP sockets -
+ * to belong to locally generated UDP traffic.
+ */
+static bool veth_skb_is_eligible_for_gro(const struct net_device *dev,
+					 const struct net_device *rcv,
+					 const struct sk_buff *skb)
+{
+	return !(dev->features & NETIF_F_ALL_TSO) ||
+		(skb->destructor == sock_wfree &&
+		 rcv->features & (NETIF_F_GRO_FRAGLIST | NETIF_F_GRO_UDP_FWD));
+}
+
 static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
 {
 	struct veth_priv *rcv_priv, *priv = netdev_priv(dev);

Currently the veth device has the GRO feature bit set, even if no GRO aggregation is possible with the default configuration, as the veth device does not hook into the GRO engine.

Flipping the GRO feature bit from user-space is a no-op, unless XDP is enabled. In such scenario GRO could actually take place, but TSO is forced to off on the peer device.

This change allow user-space to really control the GRO feature, with no need for an XDP program.

The GRO feature bit is now cleared by default - so that there are no user-visible behavior changes with the default configuration.

When the GRO bit is set, the per-queue NAPI instances are initialized and registered. On xmit, when napi instances are available, we try to use them.

Quoted 48 lines · Paolo Abeni

After the previous patch, when enabling GRO, locally generated TCP traffic experiences some measurable overhead, as it traverses the GRO engine without any chance of aggregation.

This change refine the NAPI receive path admission test, to avoid unnecessary GRO overhead in most scenarios, when GRO is enabled on a veth peer.

Only skbs that are eligible for aggregation enter the GRO layer, the others will go through the traditional receive path.

Signed-off-by Paolo Abeni <[email protected]>

---
 drivers/net/veth.c | 23 ++++++++++++++++++++++-
 1 file changed, 22 insertions(+), 1 deletion(-)

drivers/net/veth.c +19 -0

index ca44e82d1edeb..85f90f33d437e 100644
@@ -282,6 +282,25 @@ static int veth_forward_skb(struct net_device *dev, struct sk_buff *skb,
 		netif_rx(skb);
 }
 
+/* return true if the specified skb has chances of GRO aggregation
+ * Don't strive for accuracy, but try to avoid GRO overhead in the most
+ * common scenarios.
+ * When XDP is enabled, all traffic is considered eligible, as the xmit
+ * device has TSO off.
+ * When TSO is enabled on the xmit device, we are likely interested only
+ * in UDP aggregation, explicitly check for that if the skb is suspected
+ * - the sock_wfree destructor is used by UDP, ICMP and XDP sockets -
+ * to belong to locally generated UDP traffic.
+ */
+static bool veth_skb_is_eligible_for_gro(const struct net_device *dev,
+					 const struct net_device *rcv,
+					 const struct sk_buff *skb)
+{
+	return !(dev->features & NETIF_F_ALL_TSO) ||
+		(skb->destructor == sock_wfree &&
+		 rcv->features & (NETIF_F_GRO_FRAGLIST | NETIF_F_GRO_UDP_FWD));
+}
+
 static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
 {
 	struct veth_priv *rcv_priv, *priv = netdev_priv(dev);

@@ -305,8 +324,10 @@ static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
 

drivers/net/veth.c static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev) · line 332

Quoted 9 lines

drivers/net/veth.c +3 -1

 		/* The napi pointer is available when an XDP program is
 		 * attached or when GRO is enabled
+		 * Don't bother with napi/GRO if the skb can't be aggregated
 		 */
-		use_napi = rcu_access_pointer(rq->napi);
+		use_napi = rcu_access_pointer(rq->napi) &&
+			   veth_skb_is_eligible_for_gro(dev, rcv, skb);
 		skb_record_rx_queue(skb, rxq);
 	}

You just changed the 'xdp_rcv' check to this use_napi, and now you're conditioning it on GRO eligibility, so doesn't this break XDP if that was the reason NAPI was turned on in the first place?

hello,

Quoted 50 lines · Toke Høiland-Jørgensen

Quoted 48 lines · Paolo Abeni

After the previous patch, when enabling GRO, locally generated TCP traffic experiences some measurable overhead, as it traverses the GRO engine without any chance of aggregation.

This change refine the NAPI receive path admission test, to avoid unnecessary GRO overhead in most scenarios, when GRO is enabled on a veth peer.

Only skbs that are eligible for aggregation enter the GRO layer, the others will go through the traditional receive path.

Signed-off-by Paolo Abeni <[email protected]>

---
 drivers/net/veth.c | 23 ++++++++++++++++++++++-
 1 file changed, 22 insertions(+), 1 deletion(-)

drivers/net/veth.c +19 -0

index ca44e82d1edeb..85f90f33d437e 100644
@@ -282,6 +282,25 @@ static int veth_forward_skb(struct net_device *dev, struct sk_buff *skb,
 		netif_rx(skb);
 }
 
+/* return true if the specified skb has chances of GRO aggregation
+ * Don't strive for accuracy, but try to avoid GRO overhead in the most
+ * common scenarios.
+ * When XDP is enabled, all traffic is considered eligible, as the xmit
+ * device has TSO off.
+ * When TSO is enabled on the xmit device, we are likely interested only
+ * in UDP aggregation, explicitly check for that if the skb is suspected
+ * - the sock_wfree destructor is used by UDP, ICMP and XDP sockets -
+ * to belong to locally generated UDP traffic.
+ */
+static bool veth_skb_is_eligible_for_gro(const struct net_device *dev,
+					 const struct net_device *rcv,
+					 const struct sk_buff *skb)
+{
+	return !(dev->features & NETIF_F_ALL_TSO) ||
+		(skb->destructor == sock_wfree &&
+		 rcv->features & (NETIF_F_GRO_FRAGLIST | NETIF_F_GRO_UDP_FWD));
+}
+
 static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
 {
 	struct veth_priv *rcv_priv, *priv = netdev_priv(dev);

@@ -305,8 +324,10 @@ static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
 

Quoted 13 lines

drivers/net/veth.c static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev) · line 332

Quoted 9 lines

drivers/net/veth.c +3 -1

 		/* The napi pointer is available when an XDP program is
 		 * attached or when GRO is enabled
+		 * Don't bother with napi/GRO if the skb can't be aggregated
 		 */
-		use_napi = rcu_access_pointer(rq->napi);
+		use_napi = rcu_access_pointer(rq->napi) &&
+			   veth_skb_is_eligible_for_gro(dev, rcv, skb);
 		skb_record_rx_queue(skb, rxq);
 	}

You just changed the 'xdp_rcv' check to this use_napi, and now you're conditioning it on GRO eligibility, so doesn't this break XDP if that was the reason NAPI was turned on in the first place?

Thank you for the feedback.

If XDP is enabled, TSO is forced of on 'dev'
and veth_skb_is_eligible_for_gro() returns true, so napi/GRO is always used - there is no functional change when XDP is enabled.

Quoted 14 lines · Paolo Abeni

Currently the veth device has the GRO feature bit set, even if no GRO aggregation is possible with the default configuration, as the veth device does not hook into the GRO engine.

Flipping the GRO feature bit from user-space is a no-op, unless XDP is enabled. In such scenario GRO could actually take place, but TSO is forced to off on the peer device.

This change allow user-space to really control the GRO feature, with no need for an XDP program.

The GRO feature bit is now cleared by default - so that there are no user-visible behavior changes with the default configuration.

Quoted 3 lines

When the GRO bit is set, the per-queue NAPI instances are initialized and registered. On xmit, when napi instances are available, we try to use them.

Am I mistaken in thinking that this also makes XDP redirect into a veth work without having to load an XDP program on the peer device? That's been a long-outstanding thing we've been meaning to fix, so that would be awesome! :)

Quoted 16 lines · Toke Høiland-Jørgensen

Quoted 14 lines · Paolo Abeni

Currently the veth device has the GRO feature bit set, even if no GRO aggregation is possible with the default configuration, as the veth device does not hook into the GRO engine.

Flipping the GRO feature bit from user-space is a no-op, unless XDP is enabled. In such scenario GRO could actually take place, but TSO is forced to off on the peer device.

This change allow user-space to really control the GRO feature, with no need for an XDP program.

The GRO feature bit is now cleared by default - so that there are no user-visible behavior changes with the default configuration.

Quoted 8 lines

Quoted 3 lines

When the GRO bit is set, the per-queue NAPI instances are initialized and registered. On xmit, when napi instances are available, we try to use them.

Am I mistaken in thinking that this also makes XDP redirect into a veth work without having to load an XDP program on the peer device? That's been a long-outstanding thing we've been meaning to fix, so that would be awesome! :)

I have not experimented that, and I admit gross ignorance WRT this argument, but AFAICS the needed bits to get XDP redirect working on veth are the ptr_ring initialization and the napi instance available.

With this patch both are in place when GRO is enabled, so I guess XPD redirect should work, too (modulo bugs for untested scenario).

Quoted 30 lines · Paolo Abeni

Quoted 16 lines · Toke Høiland-Jørgensen

Quoted 14 lines · Paolo Abeni

Currently the veth device has the GRO feature bit set, even if no GRO aggregation is possible with the default configuration, as the veth device does not hook into the GRO engine.

Flipping the GRO feature bit from user-space is a no-op, unless XDP is enabled. In such scenario GRO could actually take place, but TSO is forced to off on the peer device.

This change allow user-space to really control the GRO feature, with no need for an XDP program.

The GRO feature bit is now cleared by default - so that there are no user-visible behavior changes with the default configuration.

Quoted 8 lines

Quoted 3 lines

When the GRO bit is set, the per-queue NAPI instances are initialized and registered. On xmit, when napi instances are available, we try to use them.

Am I mistaken in thinking that this also makes XDP redirect into a veth work without having to load an XDP program on the peer device? That's been a long-outstanding thing we've been meaning to fix, so that would be awesome! :)

I have not experimented that, and I admit gross ignorance WRT this argument, but AFAICS the needed bits to get XDP redirect working on veth are the ptr_ring initialization and the napi instance available.

Quoted 2 lines

With this patch both are in place when GRO is enabled, so I guess XPD redirect should work, too (modulo bugs for untested scenario).

OK, finally got around to testing this; it doesn't quite work with just your patch, because veth_xdp_xmit() still checks for rq->xdp_prog instead of rq->napi. Fixing this indeed enabled veth to be an XDP_REDIRECT target without an XDP program loaded on the peer. So yay! I'll send a followup fixing that check.

So with this we seem to have some nice improvements in both functionality and performance when GRO is turned on; so any reason why we shouldn't just flip the default to on?

Quoted 32 lines · Toke Høiland-Jørgensen

Quoted 30 lines · Paolo Abeni

Quoted 16 lines · Toke Høiland-Jørgensen

Quoted 14 lines · Paolo Abeni

Currently the veth device has the GRO feature bit set, even if no GRO aggregation is possible with the default configuration, as the veth device does not hook into the GRO engine.

Flipping the GRO feature bit from user-space is a no-op, unless XDP is enabled. In such scenario GRO could actually take place, but TSO is forced to off on the peer device.

This change allow user-space to really control the GRO feature, with no need for an XDP program.

The GRO feature bit is now cleared by default - so that there are no user-visible behavior changes with the default configuration.

Quoted 8 lines

Quoted 3 lines

When the GRO bit is set, the per-queue NAPI instances are initialized and registered. On xmit, when napi instances are available, we try to use them.

Am I mistaken in thinking that this also makes XDP redirect into a veth work without having to load an XDP program on the peer device? That's been a long-outstanding thing we've been meaning to fix, so that would be awesome! :)

I have not experimented that, and I admit gross ignorance WRT this argument, but AFAICS the needed bits to get XDP redirect working on veth are the ptr_ring initialization and the napi instance available.

Quoted 8 lines

Quoted 2 lines

With this patch both are in place when GRO is enabled, so I guess XPD redirect should work, too (modulo bugs for untested scenario).

OK, finally got around to testing this; it doesn't quite work with just your patch, because veth_xdp_xmit() still checks for rq->xdp_prog instead of rq->napi. Fixing this indeed enabled veth to be an XDP_REDIRECT target without an XDP program loaded on the peer. So yay! I'll send a followup fixing that check.

Thank you for double checking!

Quoted 3 lines

So with this we seem to have some nice improvements in both functionality and performance when GRO is turned on; so any reason why we shouldn't just flip the default to on?

Uhmmm... patch 3/4 should avoid the GRO overhead for most cases where we can't leverage the aggregation benefit, but I'm not 110% sure that enabling GRO by default will not cause performance regressions in some scenarios.

It this proves to be always a win we can still change the default later, I think.