Skip to contentSkip to conversation

VLAN driver question

DISCUSSION Jim Baxter 2 participants 7 messages Jun 17, 2013 – Jun 21, 2013

Subsystems (2)
lore
Share & tools
lore

Jim Baxter

author
#0 lore raw

I have a network card that has a single flag to indicate if a received packet contains a vlan packet.

Can I use this to accelerate the kernels handling of the packet with something like the following?

/* Handle received VLAN packets */
if ((ndev->features & NETIF_F_HW_VLAN_CTAG_RX) && ebdp &&

					(ebdp->cbd_esc &
					BD_ENET_RX_VLAN)) {
	/* Push and remove the vlan tag */
	struct vlan_hdr *vlan_header;
	u16 vlan_tag;
	vlan_header = (struct vlan_hdr *) skb->data;
	vlan_tag = ntohs(vlan_header->h_vlan_TCI);
	__vlan_hwaccel_put_tag(skb, vlan_tag);
	skb->len -= VLAN_HLEN;
	skb->data += VLAN_HLEN;
	vlan_set_encap_proto(skb, vlan_header);
}

napi_gro_receive(&fep->napi, skb);

Also I cannot find any documentation stating what the difference is between NETIF_F_HW_VLAN_CTAG_RX and NETIF_F_HW_VLAN_CTAG_FILTER, can anyone define it?
The FILTER word on NETIF_F_HW_VLAN_CTAG_FILTER suggests it may be applicable to received packets.

Thank you for your help and comments.
Jim Baxter

#1 lore raw
Quoted 8 lines · Jim Baxter

I have a network card that has a single flag to indicate if a received packet contains a vlan packet.

Can I use this to accelerate the kernels handling of the packet with something like the following?

/* Handle received VLAN packets */
if ((ndev->features & NETIF_F_HW_VLAN_CTAG_RX) && ebdp &&

Quoted 11 lines
					(ebdp->cbd_esc &
					BD_ENET_RX_VLAN)) {
	/* Push and remove the vlan tag */
	struct vlan_hdr *vlan_header;
	u16 vlan_tag;
	vlan_header = (struct vlan_hdr *) skb->data;
	vlan_tag = ntohs(vlan_header->h_vlan_TCI);
	__vlan_hwaccel_put_tag(skb, vlan_tag);
	skb->len -= VLAN_HLEN;
	skb->data += VLAN_HLEN;
	vlan_set_encap_proto(skb, vlan_header);

You have to move the Ethernet addresses forward as well. Imagine this device is being bridged - the Ethernet header won't be regenerated, so it has to be immediately before the network header.

Quoted 3 lines

}

napi_gro_receive(&fep->napi, skb);

I've been meaning to add a core function to do pull-out-VLAN-tag-and-GRO, although I'm more interested in napi_gro_frags(). Perhaps you could add the generic wrapper for napi_gro_receive()?

Quoted 5 lines

Also I cannot find any documentation stating what the difference is between NETIF_F_HW_VLAN_CTAG_RX and NETIF_F_HW_VLAN_CTAG_FILTER, can anyone define it?
The FILTER word on NETIF_F_HW_VLAN_CTAG_FILTER suggests it may be applicable to received packets.

NETIF_F_HW_VLAN_CTAG_RX means the device will extract VLAN tags so they aren't received inline.

NETIF_F_HW_VLAN_CTAG_FILTER means the device can filter by VID. The driver's ndo_vlan_rx_{add,kill}_vid operations will be called to update the filter.

Ben.

Signature · 3 lines
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
author
#2 lore raw

Thank you Ben, that is very helpful.

Quoted 21 lines · Ben Hutchings
Quoted 8 lines · Jim Baxter

I have a network card that has a single flag to indicate if a received packet contains a vlan packet.

Can I use this to accelerate the kernels handling of the packet with something like the following?

/* Handle received VLAN packets */
if ((ndev->features & NETIF_F_HW_VLAN_CTAG_RX) && ebdp &&

Quoted 11 lines
					(ebdp->cbd_esc &
					BD_ENET_RX_VLAN)) {
	/* Push and remove the vlan tag */
	struct vlan_hdr *vlan_header;
	u16 vlan_tag;
	vlan_header = (struct vlan_hdr *) skb->data;
	vlan_tag = ntohs(vlan_header->h_vlan_TCI);
	__vlan_hwaccel_put_tag(skb, vlan_tag);
	skb->len -= VLAN_HLEN;
	skb->data += VLAN_HLEN;
	vlan_set_encap_proto(skb, vlan_header);
Quoted 4 lines

You have to move the Ethernet addresses forward as well. Imagine this device is being bridged - the Ethernet header won't be regenerated, so it has to be immediately before the network header.

Unless I have misunderstood, the Ethernet header is already copied across to the skb by the Driver I am editing with the lines:

	skb_reserve(skb, NET_IP_ALIGN);
	skb_put(skb, pkt_len - 4);	/* Make room */
	skb_copy_to_linear_data(skb, data, pkt_len - 4);
	skb->protocol = eth_type_trans(skb, ndev);

Thinking about it though I should actually use the flag from the driver to create an skb of the size (pkt_len - 4 - VLAN_HLEN) and extract the vlan header from data so it is not copied to the skb, does this sound sensible?

Quoted 15 lines
Quoted 3 lines

}

napi_gro_receive(&fep->napi, skb);

I've been meaning to add a core function to do pull-out-VLAN-tag-and-GRO, although I'm more interested in napi_gro_frags(). Perhaps you could add the generic wrapper for napi_gro_receive()?

Quoted 5 lines

Also I cannot find any documentation stating what the difference is between NETIF_F_HW_VLAN_CTAG_RX and NETIF_F_HW_VLAN_CTAG_FILTER, can anyone define it?
The FILTER word on NETIF_F_HW_VLAN_CTAG_FILTER suggests it may be applicable to received packets.

Quoted 9 lines

NETIF_F_HW_VLAN_CTAG_RX means the device will extract VLAN tags so they aren't received inline.

NETIF_F_HW_VLAN_CTAG_FILTER means the device can filter by VID. The driver's ndo_vlan_rx_{add,kill}_vid operations will be called to update the filter.

Ben.

Jim

#3 lore raw
Quoted 28 lines · Jim Baxter

Thank you Ben, that is very helpful.

Quoted 21 lines · Ben Hutchings
Quoted 8 lines · Jim Baxter

I have a network card that has a single flag to indicate if a received packet contains a vlan packet.

Can I use this to accelerate the kernels handling of the packet with something like the following?

/* Handle received VLAN packets */
if ((ndev->features & NETIF_F_HW_VLAN_CTAG_RX) && ebdp &&

Quoted 11 lines
					(ebdp->cbd_esc &
					BD_ENET_RX_VLAN)) {
	/* Push and remove the vlan tag */
	struct vlan_hdr *vlan_header;
	u16 vlan_tag;
	vlan_header = (struct vlan_hdr *) skb->data;
	vlan_tag = ntohs(vlan_header->h_vlan_TCI);
	__vlan_hwaccel_put_tag(skb, vlan_tag);
	skb->len -= VLAN_HLEN;
	skb->data += VLAN_HLEN;
	vlan_set_encap_proto(skb, vlan_header);
Quoted 4 lines

You have to move the Ethernet addresses forward as well. Imagine this device is being bridged - the Ethernet header won't be regenerated, so it has to be immediately before the network header.

Quoted 6 lines

Unless I have misunderstood, the Ethernet header is already copied across to the skb by the Driver I am editing with the lines:

	skb_reserve(skb, NET_IP_ALIGN);
	skb_put(skb, pkt_len - 4);	/* Make room */
	skb_copy_to_linear_data(skb, data, pkt_len - 4);
	skb->protocol = eth_type_trans(skb, ndev);

Right, but the copy and ether_type_trans() result in an skb like this:

-14 -8  -2   0   2    4
|dst|src|8100|vid|type|payload...|
             ^

After extracting the VLAN tag the skb looks like:

-18 -12 -6   -4  -2   0
|dst|src|8100|vid|type|payload...|
                      ^

If this packet is bridged, the 14 bytes before skb->data will be treated as the Ethernet header:

0 2   8    10  12   14
|x|src|8100|vid|type|payload...|

^

Then with the VLAN tag inserted;

0 2   8    10  12   14  16   18
|x|src|8100|vid|8100|vid|type|payload...|
Quoted 4 lines

Thinking about it though I should actually use the flag from the driver to create an skb of the size (pkt_len - 4 - VLAN_HLEN) and extract the vlan header from data so it is not copied to the skb, does this sound sensible?

[...]

That would also work.

Ben.

Signature · 3 lines
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
author
#4 lore raw
Quoted 64 lines · Ben Hutchings
Quoted 28 lines · Jim Baxter

Thank you Ben, that is very helpful.

Quoted 21 lines · Ben Hutchings
Quoted 8 lines · Jim Baxter

I have a network card that has a single flag to indicate if a received packet contains a vlan packet.

Can I use this to accelerate the kernels handling of the packet with something like the following?

/* Handle received VLAN packets */
if ((ndev->features & NETIF_F_HW_VLAN_CTAG_RX) && ebdp &&

Quoted 11 lines
					(ebdp->cbd_esc &
					BD_ENET_RX_VLAN)) {
	/* Push and remove the vlan tag */
	struct vlan_hdr *vlan_header;
	u16 vlan_tag;
	vlan_header = (struct vlan_hdr *) skb->data;
	vlan_tag = ntohs(vlan_header->h_vlan_TCI);
	__vlan_hwaccel_put_tag(skb, vlan_tag);
	skb->len -= VLAN_HLEN;
	skb->data += VLAN_HLEN;
	vlan_set_encap_proto(skb, vlan_header);
Quoted 4 lines

You have to move the Ethernet addresses forward as well. Imagine this device is being bridged - the Ethernet header won't be regenerated, so it has to be immediately before the network header.

Quoted 6 lines

Unless I have misunderstood, the Ethernet header is already copied across to the skb by the Driver I am editing with the lines:

	skb_reserve(skb, NET_IP_ALIGN);
	skb_put(skb, pkt_len - 4);	/* Make room */
	skb_copy_to_linear_data(skb, data, pkt_len - 4);
	skb->protocol = eth_type_trans(skb, ndev);

Right, but the copy and ether_type_trans() result in an skb like this:

-14 -8  -2   0   2    4
|dst|src|8100|vid|type|payload...|
             ^

After extracting the VLAN tag the skb looks like:

-18 -12 -6   -4  -2   0
|dst|src|8100|vid|type|payload...|
                      ^

If this packet is bridged, the 14 bytes before skb->data will be treated as the Ethernet header:

0 2   8    10  12   14
|x|src|8100|vid|type|payload...|

^

Then with the VLAN tag inserted;

0 2   8    10  12   14  16   18
|x|src|8100|vid|8100|vid|type|payload...|
Quoted 4 lines

Thinking about it though I should actually use the flag from the driver to create an skb of the size (pkt_len - 4 - VLAN_HLEN) and extract the vlan header from data so it is not copied to the skb, does this sound sensible?

Quoted 6 lines

[...]

That would also work.

Ben.

Thank you for that information, I have decided to extract the data before copying the data into the skb->data structure.

This is the put of the packet showing my removal of the data correctly, however the network is not receiving any data on the VLAN.

- data buffer from the Network card
ff ff ff ff ff ff 00 26 b9 d4 19 01 81 00 00 0a 08 06 00 01 08 00 06 04 00 01 00 26 b9 d4

- skb->data after I have only copied data to it and removed the vlan 4 octet header using skb_copy_to_linear_data and skb_copy_to_linear_data_offset.
ff ff ff ff ff ff 00 26 b9 d4 19 01 08 06 00 01 08 00 06 04 00 01 00 26 b9 d4 19 01 c0 a8

- skb-data after eth_type_trans()
00 01 08 00 06 04 00 01 00 26 b9 d4 19 01 c0 a8 01 67 00 00 00 00 00 00 c0 a8 01 64 00 00

After this point I send the packet into the network stack with:

	__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);

	napi_gro_receive(&fep->napi, skb);

Am I missing a kernel call or doing something wrong to cause the packet to be rejected in the kernel?

Thank you for any help,
Jim

#5 lore raw

On Thu, 2013-06-20 at 20:05 +0100, Jim Baxter wrote:
[...]

Quoted 20 lines

Thank you for that information, I have decided to extract the data before copying the data into the skb->data structure.

This is the put of the packet showing my removal of the data correctly, however the network is not receiving any data on the VLAN.

- data buffer from the Network card
ff ff ff ff ff ff 00 26 b9 d4 19 01 81 00 00 0a 08 06 00 01 08 00 06 04 00 01 00 26 b9 d4

- skb->data after I have only copied data to it and removed the vlan 4 octet header using skb_copy_to_linear_data and skb_copy_to_linear_data_offset.
ff ff ff ff ff ff 00 26 b9 d4 19 01 08 06 00 01 08 00 06 04 00 01 00 26 b9 d4 19 01 c0 a8

- skb-data after eth_type_trans()
00 01 08 00 06 04 00 01 00 26 b9 d4 19 01 c0 a8 01 67 00 00 00 00 00 00 c0 a8 01 64 00 00

Quoted 7 lines

After this point I send the packet into the network stack with:

	__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);

	napi_gro_receive(&fep->napi, skb);

Am I missing a kernel call or doing something wrong to cause the packet to be rejected in the kernel?

That looks about right to me, but perhaps vlan_tag has the wrong byte order? __vlan_hwaccel_put_tag() expects it in host byte order.

Ben.

Signature · 3 lines
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
author
#6 lore raw
Quoted 27 lines · Ben Hutchings

On Thu, 2013-06-20 at 20:05 +0100, Jim Baxter wrote:
[...]

Quoted 25 lines

Thank you for that information, I have decided to extract the data before copying the data into the skb->data structure.

This is the put of the packet showing my removal of the data correctly, however the network is not receiving any data on the VLAN.

- data buffer from the Network card
ff ff ff ff ff ff 00 26 b9 d4 19 01 81 00 00 0a 08 06 00 01 08 00 06 04 00 01 00 26 b9 d4

- skb->data after I have only copied data to it and removed the vlan 4 octet header using skb_copy_to_linear_data and skb_copy_to_linear_data_offset.
ff ff ff ff ff ff 00 26 b9 d4 19 01 08 06 00 01 08 00 06 04 00 01 00 26 b9 d4 19 01 c0 a8

- skb-data after eth_type_trans()
00 01 08 00 06 04 00 01 00 26 b9 d4 19 01 c0 a8 01 67 00 00 00 00 00 00 c0 a8 01 64 00 00

After this point I send the packet into the network stack with:

	__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);

	napi_gro_receive(&fep->napi, skb);
Quoted 8 lines
Quoted 2 lines

Am I missing a kernel call or doing something wrong to cause the packet to be rejected in the kernel?

That looks about right to me, but perhaps vlan_tag has the wrong byte order? __vlan_hwaccel_put_tag() expects it in host byte order.

Ben.

Hi, Thank you, I have got it working. The problem was that when I had changed the code to remove the VLAN before putting the data into the skb, I was not moving the pointer past the TPID.

It now all works with ping, ssh and scp'ing small files but when I scp a large file it stalls. I suspect it is because it is not processing packets that are at the MTU of 1518 bytes (1522 bytes with the VLAN header) in size.

Jim