[RFC PATCH] virtio-net: use NETIF_F_GRO_HW instead of NETIF_F_LRO

Commit a02e8964eaf92 ("virtio-net: ethtool configurable LRO") tries to
advertise LRO on behalf of the guest offloading features and allow the
administrator to enable and disable those features via ethtool.

This may lead several issues:

- For the device that doesn't support control guest offloads, the
  "LRO" can't be disabled so we will get a warn in the
  dev_disable_lro()
- For the device that have the control guest offloads, the guest
  offloads were disabled in the case of bridge etc which may slow down
  the traffic.

Try to fix this by using NETIF_F_GRO_HW instead so we're not
guaranteed to be re-segmented as original. Or we may want a new netdev
feature like RX_GSO since the guest offloads for virtio-net is
actually to receive GSO packet.

Or we can try not advertise LRO is control guest offloads is not
enabled. This solves the warning but will still slow down the traffic.

Signed-off-by: Jason Wang <redacted>
---
 drivers/net/virtio_net.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 0416a7e00914..10c382b08bce 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -63,7 +63,7 @@ static const unsigned long guest_offloads[] = {
 	VIRTIO_NET_F_GUEST_CSUM
 };
 
-#define GUEST_OFFLOAD_LRO_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
+#define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
 				(1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
 				(1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
 				(1ULL << VIRTIO_NET_F_GUEST_UFO))

@@ -2481,7 +2481,7 @@ static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
 	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
-		NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
+		NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
 		return -EOPNOTSUPP;
 	}
 

@@ -2612,15 +2612,15 @@ static int virtnet_set_features(struct net_device *dev,
 	u64 offloads;
 	int err;
 
-	if ((dev->features ^ features) & NETIF_F_LRO) {
+	if ((dev->features ^ features) & NETIF_F_GRO_HW) {
 		if (vi->xdp_enabled)
 			return -EBUSY;
 
-		if (features & NETIF_F_LRO)
+		if (features & NETIF_F_GRO_HW)
 			offloads = vi->guest_offloads_capable;
 		else
 			offloads = vi->guest_offloads_capable &
-				   ~GUEST_OFFLOAD_LRO_MASK;
+				   ~GUEST_OFFLOAD_GRO_HW_MASK;
 
 		err = virtnet_set_guest_offloads(vi, offloads);
 		if (err)

@@ -3100,9 +3100,9 @@ static int virtnet_probe(struct virtio_device *vdev)
 		dev->features |= NETIF_F_RXCSUM;
 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
-		dev->features |= NETIF_F_LRO;
+		dev->features |= NETIF_F_GRO_HW;
 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
-		dev->hw_features |= NETIF_F_LRO;
+		dev->hw_features |= NETIF_F_GRO_HW;
 
 	dev->vlan_features = dev->features;
 

-- 
2.25.1

On Wed, 11 Aug 2021 16:16:23 +0800 Jason Wang wrote:

Try to fix this by using NETIF_F_GRO_HW instead so we're not
guaranteed to be re-segmented as original. 

This sentence may need rephrasing.

Or we may want a new netdev
feature like RX_GSO since the guest offloads for virtio-net is
actually to receive GSO packet.

Or we can try not advertise LRO is control guest offloads is not
enabled. This solves the warning but will still slow down the traffic.

IMO gro-hw fits pretty well, patch looks good.

On Wed, Aug 11, 2021 at 3:16 AM Jason Wang [off-list ref] wrote:

I applied this patch, recompiled the kernel, and tested it.
The warning messages are gone. Network speed is normal.
I can now enable forwarding, and nothing bad happens.
So far, so good.

Thank you.

在 2021/8/12 上午6:17, Jakub Kicinski 写道:

On Wed, 11 Aug 2021 16:16:23 +0800 Jason Wang wrote:

quoted

Try to fix this by using NETIF_F_GRO_HW instead so we're not
guaranteed to be re-segmented as original.

This sentence may need rephrasing.

Right, actually, I meant:


Try to fix this by using NETIF_F_GRO_HW instead. But we're not sure the 
packet could be re-segmented to the exact original packet stream. Since 
it's really depends on the bakcend implementation.

quoted

Or we may want a new netdev
feature like RX_GSO since the guest offloads for virtio-net is
actually to receive GSO packet.

Or we can try not advertise LRO is control guest offloads is not
enabled. This solves the warning but will still slow down the traffic.

IMO gro-hw fits pretty well, patch looks good.

If the re-segmentation is not a issue. I will post a formal patch.

Thanks

On Thu, Aug 12, 2021 at 11:23:04AM +0800, Jason Wang wrote:

在 2021/8/12 上午6:17, Jakub Kicinski 写道:

quoted

On Wed, 11 Aug 2021 16:16:23 +0800 Jason Wang wrote:

quoted

Try to fix this by using NETIF_F_GRO_HW instead so we're not
guaranteed to be re-segmented as original.

This sentence may need rephrasing.

Right, actually, I meant:


Try to fix this by using NETIF_F_GRO_HW instead. But we're not sure the
packet could be re-segmented to the exact original packet stream. Since it's
really depends on the bakcend implementation.

quoted

quoted

Or we may want a new netdev
feature like RX_GSO since the guest offloads for virtio-net is
actually to receive GSO packet.

Or we can try not advertise LRO is control guest offloads is not
enabled. This solves the warning but will still slow down the traffic.

IMO gro-hw fits pretty well, patch looks good.

If the re-segmentation is not a issue. I will post a formal patch.

Thanks

It is but the point is even though spec did not require this
we always allowed these configurations
in the past so hopefully most of them are not broken and combine
packets in the same way as GRO. Let's not break them all
in an attempt to catch bad configs, and meanwhile amend
the spec to recommend doing GW GRO.

quoted

On Wed, Aug 11, 2021 at 08:20:03PM -0500, ivan wrote:

On Wed, Aug 11, 2021 at 3:16 AM Jason Wang [off-list ref] wrote:

quoted

Commit a02e8964eaf92 ("virtio-net: ethtool configurable LRO") tries to
advertise LRO on behalf of the guest offloading features and allow the
administrator to enable and disable those features via ethtool.

This may lead several issues:

- For the device that doesn't support control guest offloads, the
  "LRO" can't be disabled so we will get a warn in the
  dev_disable_lro()
- For the device that have the control guest offloads, the guest
  offloads were disabled in the case of bridge etc which may slow down
  the traffic.

Try to fix this by using NETIF_F_GRO_HW instead so we're not
guaranteed to be re-segmented as original. Or we may want a new netdev
feature like RX_GSO since the guest offloads for virtio-net is
actually to receive GSO packet.

Or we can try not advertise LRO is control guest offloads is not
enabled. This solves the warning but will still slow down the traffic.

Signed-off-by: Jason Wang <redacted>
---
 drivers/net/virtio_net.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 0416a7e00914..10c382b08bce 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -63,7 +63,7 @@ static const unsigned long guest_offloads[] = {
        VIRTIO_NET_F_GUEST_CSUM
 };

-#define GUEST_OFFLOAD_LRO_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
+#define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
                                (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
                                (1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
                                (1ULL << VIRTIO_NET_F_GUEST_UFO))

@@ -2481,7 +2481,7 @@ static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
                virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
                virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
                virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
-               NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
+               NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
                return -EOPNOTSUPP;
        }

@@ -2612,15 +2612,15 @@ static int virtnet_set_features(struct net_device *dev,
        u64 offloads;
        int err;

-       if ((dev->features ^ features) & NETIF_F_LRO) {
+       if ((dev->features ^ features) & NETIF_F_GRO_HW) {
                if (vi->xdp_enabled)
                        return -EBUSY;

-               if (features & NETIF_F_LRO)
+               if (features & NETIF_F_GRO_HW)
                        offloads = vi->guest_offloads_capable;
                else
                        offloads = vi->guest_offloads_capable &
-                                  ~GUEST_OFFLOAD_LRO_MASK;
+                                  ~GUEST_OFFLOAD_GRO_HW_MASK;

                err = virtnet_set_guest_offloads(vi, offloads);
                if (err)

@@ -3100,9 +3100,9 @@ static int virtnet_probe(struct virtio_device *vdev)
                dev->features |= NETIF_F_RXCSUM;
        if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
            virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
-               dev->features |= NETIF_F_LRO;
+               dev->features |= NETIF_F_GRO_HW;
        if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
-               dev->hw_features |= NETIF_F_LRO;
+               dev->hw_features |= NETIF_F_GRO_HW;

        dev->vlan_features = dev->features;

--

I applied this patch, recompiled the kernel, and tested it.
The warning messages are gone. Network speed is normal.
I can now enable forwarding, and nothing bad happens.
So far, so good.

Thank you.

OK so that's

Tested-by: ivan <redacted>

It is still weird that without the patch networking dies.

What happens if you apply the patch then try to disable GRO
using ethtool?

-- 
MST

在 2021/8/12 下午12:50, Michael S. Tsirkin 写道:

On Thu, Aug 12, 2021 at 11:23:04AM +0800, Jason Wang wrote:

quoted

在 2021/8/12 上午6:17, Jakub Kicinski 写道:

quoted

On Wed, 11 Aug 2021 16:16:23 +0800 Jason Wang wrote:

quoted

Try to fix this by using NETIF_F_GRO_HW instead so we're not
guaranteed to be re-segmented as original.

This sentence may need rephrasing.

Right, actually, I meant:


Try to fix this by using NETIF_F_GRO_HW instead. But we're not sure the
packet could be re-segmented to the exact original packet stream. Since it's
really depends on the bakcend implementation.

quoted

quoted

Or we may want a new netdev
feature like RX_GSO since the guest offloads for virtio-net is
actually to receive GSO packet.

Or we can try not advertise LRO is control guest offloads is not
enabled. This solves the warning but will still slow down the traffic.

IMO gro-hw fits pretty well, patch looks good.

If the re-segmentation is not a issue. I will post a formal patch.

Thanks

It is but the point is even though spec did not require this
we always allowed these configurations
in the past so hopefully most of them are not broken and combine
packets in the same way as GRO. Let's not break them all
in an attempt to catch bad configs, and meanwhile amend
the spec to recommend doing GW GRO.

Ok, let me add this in the commit log and send a formal patch.

Thanks

On Thu, Aug 12, 2021 at 12:00 AM Michael S. Tsirkin [off-list ref] wrote:

On Wed, Aug 11, 2021 at 08:20:03PM -0500, ivan wrote:

quoted

On Wed, Aug 11, 2021 at 3:16 AM Jason Wang [off-list ref] wrote:

quoted

Commit a02e8964eaf92 ("virtio-net: ethtool configurable LRO") tries to
advertise LRO on behalf of the guest offloading features and allow the
administrator to enable and disable those features via ethtool.

This may lead several issues:

- For the device that doesn't support control guest offloads, the
  "LRO" can't be disabled so we will get a warn in the
  dev_disable_lro()
- For the device that have the control guest offloads, the guest
  offloads were disabled in the case of bridge etc which may slow down
  the traffic.

Try to fix this by using NETIF_F_GRO_HW instead so we're not
guaranteed to be re-segmented as original. Or we may want a new netdev
feature like RX_GSO since the guest offloads for virtio-net is
actually to receive GSO packet.

Or we can try not advertise LRO is control guest offloads is not
enabled. This solves the warning but will still slow down the traffic.

Signed-off-by: Jason Wang <redacted>
---
 drivers/net/virtio_net.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 0416a7e00914..10c382b08bce 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -63,7 +63,7 @@ static const unsigned long guest_offloads[] = {
        VIRTIO_NET_F_GUEST_CSUM
 };

-#define GUEST_OFFLOAD_LRO_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
+#define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
                                (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
                                (1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
                                (1ULL << VIRTIO_NET_F_GUEST_UFO))

@@ -2481,7 +2481,7 @@ static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
                virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
                virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
                virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
-               NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
+               NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
                return -EOPNOTSUPP;
        }

@@ -2612,15 +2612,15 @@ static int virtnet_set_features(struct net_device *dev,
        u64 offloads;
        int err;

-       if ((dev->features ^ features) & NETIF_F_LRO) {
+       if ((dev->features ^ features) & NETIF_F_GRO_HW) {
                if (vi->xdp_enabled)
                        return -EBUSY;

-               if (features & NETIF_F_LRO)
+               if (features & NETIF_F_GRO_HW)
                        offloads = vi->guest_offloads_capable;
                else
                        offloads = vi->guest_offloads_capable &
-                                  ~GUEST_OFFLOAD_LRO_MASK;
+                                  ~GUEST_OFFLOAD_GRO_HW_MASK;

                err = virtnet_set_guest_offloads(vi, offloads);
                if (err)

@@ -3100,9 +3100,9 @@ static int virtnet_probe(struct virtio_device *vdev)
                dev->features |= NETIF_F_RXCSUM;
        if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
            virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
-               dev->features |= NETIF_F_LRO;
+               dev->features |= NETIF_F_GRO_HW;
        if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
-               dev->hw_features |= NETIF_F_LRO;
+               dev->hw_features |= NETIF_F_GRO_HW;

        dev->vlan_features = dev->features;

--

I applied this patch, recompiled the kernel, and tested it.
The warning messages are gone. Network speed is normal.
I can now enable forwarding, and nothing bad happens.
So far, so good.

Thank you.

OK so that's

Tested-by: ivan <redacted>

It is still weird that without the patch networking dies.

What happens if you apply the patch then try to disable GRO
using ethtool?

Nothing bad.  Ethtool shows successful change to off.
Makes no differrece on the iperf tests.  Still good.