Currently, netpoll only supports interfaces with an ethernet-compatible
link layer. Certain interfaces like SLIP do not have a link layer
on the network interface level at all and expect raw IP packets,
and could benefit from being supported by netpoll.

This commit adds support for such interfaces by using the network device's
`hard_header_len` field as an indication that no link layer is present.
If that is the case we simply skip adding the ethernet header, causing
a raw IP packet to be sent over the interface. This has been confirmed
to add netconsole support to at least SLIP and WireGuard interfaces.

Signed-off-by: Mark Cilissen <mark@yotsuba.nl>
---
 Documentation/networking/netconsole.rst |  3 ++-
 net/core/netpoll.c                      | 30 ++++++++++++++++---------
 2 files changed, 22 insertions(+), 11 deletions(-)

diff --git a/Documentation/networking/netconsole.rst b/Documentation/networking/netconsole.rst
index d55c2a22ec7a..434ce0366027 100644
--- a/Documentation/networking/netconsole.rst
+++ b/Documentation/networking/netconsole.rst
@@ -327,4 +327,5 @@ enable the logging of even the most critical kernel bugs. It works
 from IRQ contexts as well, and does not enable interrupts while
 sending packets. Due to these unique needs, configuration cannot
 be more automatic, and some fundamental limitations will remain:
-only IP networks, UDP packets and ethernet devices are supported.
+only UDP packets over IP networks, over ethernet links if a
+hardware layer is required, are supported.
diff --git a/net/core/netpoll.c b/net/core/netpoll.c
index 543007f159f9..0299fb71b456 100644
--- a/net/core/netpoll.c
+++ b/net/core/netpoll.c
@@ -399,7 +399,7 @@ EXPORT_SYMBOL(netpoll_send_skb);
 
 void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
 {
-	int total_len, ip_len, udp_len;
+	int total_len, ip_len, udp_len, link_len;
 	struct sk_buff *skb;
 	struct udphdr *udph;
 	struct iphdr *iph;
@@ -416,7 +416,10 @@ void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
 	else
 		ip_len = udp_len + sizeof(*iph);
 
-	total_len = ip_len + LL_RESERVED_SPACE(np->dev);
+	/* if there's a hardware header assume ethernet, else raw IP */
+	eth = NULL;
+	link_len = np->dev->hard_header_len ? LL_RESERVED_SPACE(np->dev) : 0;
+	total_len = ip_len + link_len;
 
 	skb = find_skb(np, total_len + np->dev->needed_tailroom,
 		       total_len - len);
@@ -458,9 +461,11 @@ void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
 		ip6h->saddr = np->local_ip.in6;
 		ip6h->daddr = np->remote_ip.in6;
 
-		eth = skb_push(skb, ETH_HLEN);
-		skb_reset_mac_header(skb);
-		skb->protocol = eth->h_proto = htons(ETH_P_IPV6);
+		skb->protocol = htons(ETH_P_IPV6);
+		if (link_len) {
+			eth = skb_push(skb, ETH_HLEN);
+			skb_reset_mac_header(skb);
+		}
 	} else {
 		udph->check = 0;
 		udph->check = csum_tcpudp_magic(np->local_ip.ip,
@@ -487,13 +492,18 @@ void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
 		put_unaligned(np->remote_ip.ip, &(iph->daddr));
 		iph->check    = ip_fast_csum((unsigned char *)iph, iph->ihl);
 
-		eth = skb_push(skb, ETH_HLEN);
-		skb_reset_mac_header(skb);
-		skb->protocol = eth->h_proto = htons(ETH_P_IP);
+		skb->protocol = htons(ETH_P_IP);
+		if (link_len) {
+			eth = skb_push(skb, ETH_HLEN);
+			skb_reset_mac_header(skb);
+		}
 	}
 
-	ether_addr_copy(eth->h_source, np->dev->dev_addr);
-	ether_addr_copy(eth->h_dest, np->remote_mac);
+	if (eth) {
+		eth->h_proto = skb->protocol;
+		ether_addr_copy(eth->h_source, np->dev->dev_addr);
+		ether_addr_copy(eth->h_dest, np->remote_mac);
+	}
 
 	skb->dev = np->dev;
 
-- 
2.43.0

On Wed, 13 Mar 2024 13:46:13 +0100 Mark Cilissen wrote:
> Currently, netpoll only supports interfaces with an ethernet-compatible
> link layer. Certain interfaces like SLIP do not have a link layer
> on the network interface level at all and expect raw IP packets,
> and could benefit from being supported by netpoll.
> 
> This commit adds support for such interfaces by using the network device's
> `hard_header_len` field as an indication that no link layer is present.
> If that is the case we simply skip adding the ethernet header, causing
> a raw IP packet to be sent over the interface. This has been confirmed
> to add netconsole support to at least SLIP and WireGuard interfaces.

Would be great if this could come with a simple selftest under
tools/testing/selftests/net. Preferably using some simple tunnel
device, rather than wg to limit tooling dependencies ;)

Hi Jakub,

> Op 14 mrt 6 Reiwa, om 19:34 heeft Jakub Kicinski <kuba@kernel.org> het volgende geschreven:
> 
> On Wed, 13 Mar 2024 13:46:13 +0100 Mark Cilissen wrote:
>> Currently, netpoll only supports interfaces with an ethernet-compatible
>> link layer. Certain interfaces like SLIP do not have a link layer
>> on the network interface level at all and expect raw IP packets,
>> and could benefit from being supported by netpoll.
>> 
>> This commit adds support for such interfaces by using the network device's
>> `hard_header_len` field as an indication that no link layer is present.
>> If that is the case we simply skip adding the ethernet header, causing
>> a raw IP packet to be sent over the interface. This has been confirmed
>> to add netconsole support to at least SLIP and WireGuard interfaces.
> 
> Would be great if this could come with a simple selftest under
> tools/testing/selftests/net. Preferably using some simple tunnel
> device, rather than wg to limit tooling dependencies ;)

Yes, that makes a lot of sense. I wrote a selftest that tests netconsole
using IPIP and GRE tunnels, and they too seem to work correctly and pass (and
more importantly, fail without the patch ;-). Would you prefer me to submit
a v2 with that test now, or when net-next reopens in a week?

Thanks and regards,
Mark

On Mon, 18 Mar 2024 12:47:46 +0100 Mark wrote:
> > Would be great if this could come with a simple selftest under
> > tools/testing/selftests/net. Preferably using some simple tunnel
> > device, rather than wg to limit tooling dependencies ;)  
> 
> Yes, that makes a lot of sense. I wrote a selftest that tests netconsole
> using IPIP and GRE tunnels, and they too seem to work correctly and pass (and
> more importantly, fail without the patch ;-). Would you prefer me to submit
> a v2 with that test now, or when net-next reopens in a week?

Feel free to post as [RFC net-next v2] now. I can't guarantee people
will find the time to review, but it shouldn't hurt to have it on 
the list :)

On Wed, 2024-03-13 at 13:46 +0100, Mark Cilissen wrote:
> Currently, netpoll only supports interfaces with an ethernet-compatible
> link layer. Certain interfaces like SLIP do not have a link layer
> on the network interface level at all and expect raw IP packets,
> and could benefit from being supported by netpoll.
> 
> This commit adds support for such interfaces by using the network device's
> `hard_header_len` field as an indication that no link layer is present.
> If that is the case we simply skip adding the ethernet header, causing
> a raw IP packet to be sent over the interface. This has been confirmed
> to add netconsole support to at least SLIP and WireGuard interfaces.
> 
> Signed-off-by: Mark Cilissen <mark@yotsuba.nl>

## Form letter - net-next-closed

The merge window for v6.9 has begun and we have already posted our pull
request. Therefore net-next is closed for new drivers, features, code
refactoring and optimizations. We are currently accepting bug fixes
only.

Please repost when net-next reopens after March 25th.

RFC patches sent for review only are obviously welcome at any time.

See:
https://www.kernel.org/doc/html/next/process/maintainer-netdev.html#development-cycle

--
pw-bot: defer

On 2024-03-13 at 18:16:13, Mark Cilissen (mark@yotsuba.nl) wrote:
> Currently, netpoll only supports interfaces with an ethernet-compatible
> link layer. Certain interfaces like SLIP do not have a link layer
> on the network interface level at all and expect raw IP packets,
> and could benefit from being supported by netpoll.
>
> This commit adds support for such interfaces by using the network device's
> `hard_header_len` field as an indication that no link layer is present.
> If that is the case we simply skip adding the ethernet header, causing
> a raw IP packet to be sent over the interface. This has been confirmed
> to add netconsole support to at least SLIP and WireGuard interfaces.
>
> Signed-off-by: Mark Cilissen <mark@yotsuba.nl>
> ---
>  Documentation/networking/netconsole.rst |  3 ++-
>  net/core/netpoll.c                      | 30 ++++++++++++++++---------
>  2 files changed, 22 insertions(+), 11 deletions(-)
>
> diff --git a/Documentation/networking/netconsole.rst b/Documentation/networking/netconsole.rst
> index d55c2a22ec7a..434ce0366027 100644
> --- a/Documentation/networking/netconsole.rst
> +++ b/Documentation/networking/netconsole.rst
> @@ -327,4 +327,5 @@ enable the logging of even the most critical kernel bugs. It works
>  from IRQ contexts as well, and does not enable interrupts while
>  sending packets. Due to these unique needs, configuration cannot
>  be more automatic, and some fundamental limitations will remain:
> -only IP networks, UDP packets and ethernet devices are supported.
> +only UDP packets over IP networks, over ethernet links if a
> +hardware layer is required, are supported.
> diff --git a/net/core/netpoll.c b/net/core/netpoll.c
> index 543007f159f9..0299fb71b456 100644
> --- a/net/core/netpoll.c
> +++ b/net/core/netpoll.c
> @@ -399,7 +399,7 @@ EXPORT_SYMBOL(netpoll_send_skb);
>
>  void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
>  {
> -	int total_len, ip_len, udp_len;
> +	int total_len, ip_len, udp_len, link_len;
>  	struct sk_buff *skb;
>  	struct udphdr *udph;
>  	struct iphdr *iph;
> @@ -416,7 +416,10 @@ void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
>  	else
>  		ip_len = udp_len + sizeof(*iph);
>
> -	total_len = ip_len + LL_RESERVED_SPACE(np->dev);
> +	/* if there's a hardware header assume ethernet, else raw IP */
Taking an assumption based on dev's lower layer does not look to be good.
why not transmit packet from skb_network_header() in your driver (by making
changes in your driver)

> +	eth = NULL;
> +	link_len = np->dev->hard_header_len ? LL_RESERVED_SPACE(np->dev) : 0;
> +	total_len = ip_len + link_len;
>
>  	skb = find_skb(np, total_len + np->dev->needed_tailroom,
>  		       total_len - len);
> @@ -458,9 +461,11 @@ void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
>  		ip6h->saddr = np->local_ip.in6;
>  		ip6h->daddr = np->remote_ip.in6;
>
> -		eth = skb_push(skb, ETH_HLEN);
> -		skb_reset_mac_header(skb);
> -		skb->protocol = eth->h_proto = htons(ETH_P_IPV6);
> +		skb->protocol = htons(ETH_P_IPV6);
> +		if (link_len) {
> +			eth = skb_push(skb, ETH_HLEN);
> +			skb_reset_mac_header(skb);
> +		}
>  	} else {
>  		udph->check = 0;
>  		udph->check = csum_tcpudp_magic(np->local_ip.ip,
> @@ -487,13 +492,18 @@ void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
>  		put_unaligned(np->remote_ip.ip, &(iph->daddr));
>  		iph->check    = ip_fast_csum((unsigned char *)iph, iph->ihl);
>
> -		eth = skb_push(skb, ETH_HLEN);
> -		skb_reset_mac_header(skb);
> -		skb->protocol = eth->h_proto = htons(ETH_P_IP);
> +		skb->protocol = htons(ETH_P_IP);
> +		if (link_len) {
> +			eth = skb_push(skb, ETH_HLEN);
> +			skb_reset_mac_header(skb);
> +		}
>  	}
>
> -	ether_addr_copy(eth->h_source, np->dev->dev_addr);
> -	ether_addr_copy(eth->h_dest, np->remote_mac);
> +	if (eth) {
> +		eth->h_proto = skb->protocol;
> +		ether_addr_copy(eth->h_source, np->dev->dev_addr);
> +		ether_addr_copy(eth->h_dest, np->remote_mac);
> +	}
>
>  	skb->dev = np->dev;
>
> --
> 2.43.0
>

Hi Ratheesh,

> Op 13 mrt 6 Reiwa, om 14:36 heeft Ratheesh Kannoth <rkannoth@marvell.com> het volgende geschreven:
> 
> On 2024-03-13 at 18:16:13, Mark Cilissen (mark@yotsuba.nl) wrote:
>> […]
> Taking an assumption based on dev’s lower layer does not look to be good.
> why not transmit packet from skb_network_header() in your driver (by making
> changes in your driver)

There’s two assumptions at play here:
- The lower layer is ethernet: this has always been present in netpoll, and is even
  documented in netconsole.rst. This comment just mentions it because we add a way
  to bypass the assumption; it is not an assumption this patch adds to the code.
- hard_header_len==0 means that there is no exposed link layer: this is a rather
  conservative assumption in my opinion, and is also mentioned in the definition
  of LL_RESERVED_SPACE:

> * Alternative is:
> *   dev->hard_header_len ? (dev->hard_header_len +
> *                            (HH_DATA_MOD - 1)) & ~(HH_DATA_MOD - 1) : 0

  The same assumption is also made in more places in the core network code, like af_packet:

>   - If the device has no dev->header_ops->create, there is no LL header
>     visible above the device. In this case, its hard_header_len should be 0.
>     The device may prepend its own header internally. In this case, its
>     needed_headroom should be set to the space needed for it to add its
>     internal header.

I could change it to, like af_packet, check `dev->header_ops` instead if that is preferred,
but I don’t think that patching every single raw IP driver to deal with skbs that managed
to somehow have link layer data would be preferred here, especially since netpoll is kind
of a special case to begin with. I am open to suggestions and ideas, though.

> […]

Thanks and regards,
Mark