[PATCH net-next] net: ipv4: Add ability to have GRE ignore DF bit in IPv4 payloads

From: Philip Prindeville <redacted>

In the presence of firewalls which improperly block ICMP Unreachable
(including Fragmentation Required) messages, Path MTU Discovery is
prevented from working.

A workaround is to handle IPv4 payloads opaquely, ignoring the DF bit--as
is done for other payloads like AppleTalk--and doing transparent
fragmentation and reassembly.

Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Philip Prindeville <redacted>
---
 include/net/ip_tunnels.h       |    1 +
 include/uapi/linux/if_tunnel.h |    1 +
 net/ipv4/ip_gre.c              |   13 +++++++++++--
 net/ipv4/ip_tunnel.c           |    2 +-
 4 files changed, 14 insertions(+), 3 deletions(-)

diff --git a/include/net/ip_tunnels.h b/include/net/ip_tunnels.h
index dbf4444..9222678 100644
--- a/include/net/ip_tunnels.h
+++ b/include/net/ip_tunnels.h

@@ -132,6 +132,7 @@ struct ip_tunnel {
 	int			ip_tnl_net_id;
 	struct gro_cells	gro_cells;
 	bool			collect_md;
+	bool			ignore_df;
 };
 
 #define TUNNEL_CSUM		__cpu_to_be16(0x01)

diff --git a/include/uapi/linux/if_tunnel.h b/include/uapi/linux/if_tunnel.h
index af4de90..1046f55 100644
--- a/include/uapi/linux/if_tunnel.h
+++ b/include/uapi/linux/if_tunnel.h

@@ -113,6 +113,7 @@ enum {
 	IFLA_GRE_ENCAP_SPORT,
 	IFLA_GRE_ENCAP_DPORT,
 	IFLA_GRE_COLLECT_METADATA,
+	IFLA_GRE_IGNORE_DF,
 	__IFLA_GRE_MAX,
 };
 

diff --git a/net/ipv4/ip_gre.c b/net/ipv4/ip_gre.c
index 4d2025f..cbeecfb 100644
--- a/net/ipv4/ip_gre.c
+++ b/net/ipv4/ip_gre.c

@@ -846,6 +846,8 @@ static void ipgre_netlink_parms(struct net_device *dev,
 				struct nlattr *tb[],
 				struct ip_tunnel_parm *parms)
 {
+	struct ip_tunnel *t = netdev_priv(dev);
+
 	memset(parms, 0, sizeof(*parms));
 
 	parms->iph.protocol = IPPROTO_GRE;

@@ -884,12 +886,13 @@ static void ipgre_netlink_parms(struct net_device *dev,
 		parms->iph.frag_off = htons(IP_DF);
 
 	if (data[IFLA_GRE_COLLECT_METADATA]) {
-		struct ip_tunnel *t = netdev_priv(dev);
-
 		t->collect_md = true;
 		if (dev->type == ARPHRD_IPGRE)
 			dev->type = ARPHRD_NONE;
 	}
+
+	if (data[IFLA_GRE_IGNORE_DF])
+		t->ignore_df = !!nla_get_u8(data[IFLA_GRE_IGNORE_DF]);
 }
 
 /* This function returns true when ENCAP attributes are present in the nl msg */

@@ -1024,6 +1027,8 @@ static size_t ipgre_get_size(const struct net_device *dev)
 		nla_total_size(2) +
 		/* IFLA_GRE_COLLECT_METADATA */
 		nla_total_size(0) +
+		/* IFLA_GRE_IGNORE_DF */
+		nla_total_size(1) +
 		0;
 }
 

@@ -1057,6 +1062,9 @@ static int ipgre_fill_info(struct sk_buff *skb, const struct net_device *dev)
 			t->encap.flags))
 		goto nla_put_failure;
 
+	if (nla_put_u8(skb, IFLA_GRE_IGNORE_DF, t->ignore_df))
+		goto nla_put_failure;
+
 	if (t->collect_md) {
 		if (nla_put_flag(skb, IFLA_GRE_COLLECT_METADATA))
 			goto nla_put_failure;

@@ -1084,6 +1092,7 @@ static const struct nla_policy ipgre_policy[IFLA_GRE_MAX + 1] = {
 	[IFLA_GRE_ENCAP_SPORT]	= { .type = NLA_U16 },
 	[IFLA_GRE_ENCAP_DPORT]	= { .type = NLA_U16 },
 	[IFLA_GRE_COLLECT_METADATA]	= { .type = NLA_FLAG },
+	[IFLA_GRE_IGNORE_DF]	= { .type = NLA_U8 },
 };
 
 static struct rtnl_link_ops ipgre_link_ops __read_mostly = {

diff --git a/net/ipv4/ip_tunnel.c b/net/ipv4/ip_tunnel.c
index d8f5e0a..95649eb 100644
--- a/net/ipv4/ip_tunnel.c
+++ b/net/ipv4/ip_tunnel.c

@@ -682,7 +682,7 @@ void ip_tunnel_xmit(struct sk_buff *skb, struct net_device *dev,
 	}
 
 	df = tnl_params->frag_off;
-	if (skb->protocol == htons(ETH_P_IP))
+	if (skb->protocol == htons(ETH_P_IP) && !tunnel->ignore_df)
 		df |= (inner_iph->frag_off&htons(IP_DF));
 
 	max_headroom = LL_RESERVED_SPACE(rt->dst.dev) + sizeof(struct iphdr)

On Tue, Jun 7, 2016, at 21:48, Philip Prindeville wrote:

Is still enforce that we inherit the ttl from the inner header to the
outer header if we don't set df bit?

Thanks,
Hannes

On 06/07/2016 02:23 PM, Hannes Frederic Sowa wrote:

On Tue, Jun 7, 2016, at 21:48, Philip Prindeville wrote:

quoted

  static struct rtnl_link_ops ipgre_link_ops __read_mostly = {

diff --git a/net/ipv4/ip_tunnel.c b/net/ipv4/ip_tunnel.c
index d8f5e0a..95649eb 100644
--- a/net/ipv4/ip_tunnel.c
+++ b/net/ipv4/ip_tunnel.c

@@ -682,7 +682,7 @@ void ip_tunnel_xmit(struct sk_buff *skb, struct

net_device *dev,
  	}
  
  	df = tnl_params->frag_off;
-       if (skb->protocol == htons(ETH_P_IP))
+       if (skb->protocol == htons(ETH_P_IP) && !tunnel->ignore_df)
  		df |= (inner_iph->frag_off&htons(IP_DF));
  
  	max_headroom = LL_RESERVED_SPACE(rt->dst.dev) + sizeof(struct iphdr)

Is still enforce that we inherit the ttl from the inner header to the
outer header if we don't set df bit?

Thanks,
Hannes

That's correct.  That happens a few lines above and we don't change that 
behavior with this patch.

-Philip

On 07.06.2016 22:42, Philip Prindeville wrote:

On 06/07/2016 02:23 PM, Hannes Frederic Sowa wrote:

quoted

On Tue, Jun 7, 2016, at 21:48, Philip Prindeville wrote:

quoted

  static struct rtnl_link_ops ipgre_link_ops __read_mostly = {

diff --git a/net/ipv4/ip_tunnel.c b/net/ipv4/ip_tunnel.c
index d8f5e0a..95649eb 100644
--- a/net/ipv4/ip_tunnel.c
+++ b/net/ipv4/ip_tunnel.c

@@ -682,7 +682,7 @@ void ip_tunnel_xmit(struct sk_buff *skb, struct

net_device *dev,
      }
        df = tnl_params->frag_off;
-       if (skb->protocol == htons(ETH_P_IP))
+       if (skb->protocol == htons(ETH_P_IP) && !tunnel->ignore_df)
          df |= (inner_iph->frag_off&htons(IP_DF));
        max_headroom = LL_RESERVED_SPACE(rt->dst.dev) + sizeof(struct
iphdr)

Is still enforce that we inherit the ttl from the inner header to the
outer header if we don't set df bit?

Thanks,
Hannes

That's correct.  That happens a few lines above and we don't change that
behavior with this patch.

It should be discouraged to use this feature, but I have to admit that I
already used stripping of DF bit, too.

Acked-by: Hannes Frederic Sowa <redacted>

Thanks,
Hannes

On 06/07/2016 04:23 PM, Hannes Frederic Sowa wrote:

On 07.06.2016 22:42, Philip Prindeville wrote:

quoted

On 06/07/2016 02:23 PM, Hannes Frederic Sowa wrote:

quoted

On Tue, Jun 7, 2016, at 21:48, Philip Prindeville wrote:

quoted

   static struct rtnl_link_ops ipgre_link_ops __read_mostly = {

diff --git a/net/ipv4/ip_tunnel.c b/net/ipv4/ip_tunnel.c
index d8f5e0a..95649eb 100644
--- a/net/ipv4/ip_tunnel.c
+++ b/net/ipv4/ip_tunnel.c

@@ -682,7 +682,7 @@ void ip_tunnel_xmit(struct sk_buff *skb, struct

net_device *dev,
       }
         df = tnl_params->frag_off;
-       if (skb->protocol == htons(ETH_P_IP))
+       if (skb->protocol == htons(ETH_P_IP) && !tunnel->ignore_df)
           df |= (inner_iph->frag_off&htons(IP_DF));
         max_headroom = LL_RESERVED_SPACE(rt->dst.dev) + sizeof(struct
iphdr)

Is still enforce that we inherit the ttl from the inner header to the
outer header if we don't set df bit?

Thanks,
Hannes

That's correct.  That happens a few lines above and we don't change that
behavior with this patch.

It should be discouraged to use this feature, but I have to admit that I
already used stripping of DF bit, too.

Acked-by: Hannes Frederic Sowa <redacted>

Thanks,
Hannes

Yeah, it's admittedly a kludge to get around broken customer firewalls 
which are discarding ICMP's they have no business filtering.

It's the "least ugly workaround" in this scenario.

If you have control over the offending firewalls, it's easier (and 
better) to fix them than configuring this option... but sometimes you 
have to put up with other people having broken equipment on your network.

-Philip

On Tue, Jun 7, 2016 at 12:48 PM, Philip Prindeville
[off-list ref] wrote:

Why are you passing this as a u8?  Why not just pass it as a flag
since it defaults to off?

Switching it to a flag would help reduce the total size.

You could just copy the approach used for COLLECT_METADATA and skip
the need to pass any data and instead just use a flag.

Looking at this it appears that you should probably make the feature
mutually exclusive from the pmtudisc feature.  It wouldn't make much
sense to set the tunnel to ignore the inner DF bit and then turn
around and set it because you are enforcing path MTU discovery on the
tunnel itself.

On Jun 7, 2016, at 9:22 PM, Alexander Duyck [off-list ref] wrote:

On Tue, Jun 7, 2016 at 12:48 PM, Philip Prindeville
[off-list ref] wrote:

quoted

From: Philip Prindeville <redacted>

In the presence of firewalls which improperly block ICMP Unreachable
(including Fragmentation Required) messages, Path MTU Discovery is
prevented from working.

A workaround is to handle IPv4 payloads opaquely, ignoring the DF bit--as
is done for other payloads like AppleTalk--and doing transparent
fragmentation and reassembly.

Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Philip Prindeville <redacted>
---
include/net/ip_tunnels.h       |    1 +
include/uapi/linux/if_tunnel.h |    1 +
net/ipv4/ip_gre.c              |   13 +++++++++++--
net/ipv4/ip_tunnel.c           |    2 +-
4 files changed, 14 insertions(+), 3 deletions(-)

diff --git a/include/net/ip_tunnels.h b/include/net/ip_tunnels.h
index dbf4444..9222678 100644
--- a/include/net/ip_tunnels.h
+++ b/include/net/ip_tunnels.h

@@ -132,6 +132,7 @@ struct ip_tunnel {
       int                     ip_tnl_net_id;
       struct gro_cells        gro_cells;
       bool                    collect_md;
+       bool                    ignore_df;

};

#define TUNNEL_CSUM            __cpu_to_be16(0x01)

diff --git a/include/uapi/linux/if_tunnel.h b/include/uapi/linux/if_tunnel.h
index af4de90..1046f55 100644
--- a/include/uapi/linux/if_tunnel.h
+++ b/include/uapi/linux/if_tunnel.h

@@ -113,6 +113,7 @@ enum {
       IFLA_GRE_ENCAP_SPORT,
       IFLA_GRE_ENCAP_DPORT,
       IFLA_GRE_COLLECT_METADATA,
+       IFLA_GRE_IGNORE_DF,
       __IFLA_GRE_MAX,

};

diff --git a/net/ipv4/ip_gre.c b/net/ipv4/ip_gre.c
index 4d2025f..cbeecfb 100644
--- a/net/ipv4/ip_gre.c
+++ b/net/ipv4/ip_gre.c

@@ -846,6 +846,8 @@ static void ipgre_netlink_parms(struct net_device *dev,
                               struct nlattr *tb[],
                               struct ip_tunnel_parm *parms)

{
+       struct ip_tunnel *t = netdev_priv(dev);
+
       memset(parms, 0, sizeof(*parms));

       parms->iph.protocol = IPPROTO_GRE;

@@ -884,12 +886,13 @@ static void ipgre_netlink_parms(struct net_device *dev,
               parms->iph.frag_off = htons(IP_DF);

       if (data[IFLA_GRE_COLLECT_METADATA]) {
-               struct ip_tunnel *t = netdev_priv(dev);
-
               t->collect_md = true;
               if (dev->type == ARPHRD_IPGRE)
                       dev->type = ARPHRD_NONE;
       }
+
+       if (data[IFLA_GRE_IGNORE_DF])
+               t->ignore_df = !!nla_get_u8(data[IFLA_GRE_IGNORE_DF]);

Why are you passing this as a u8?  Why not just pass it as a flag
since it defaults to off?

quoted

}

/* This function returns true when ENCAP attributes are present in the nl msg */

@@ -1024,6 +1027,8 @@ static size_t ipgre_get_size(const struct net_device *dev)
               nla_total_size(2) +
               /* IFLA_GRE_COLLECT_METADATA */
               nla_total_size(0) +
+               /* IFLA_GRE_IGNORE_DF */
+               nla_total_size(1) +
               0;

}

Switching it to a flag would help reduce the total size.

quoted

@@ -1057,6 +1062,9 @@ static int ipgre_fill_info(struct sk_buff *skb, const struct net_device *dev)
                       t->encap.flags))
               goto nla_put_failure;

+       if (nla_put_u8(skb, IFLA_GRE_IGNORE_DF, t->ignore_df))
+               goto nla_put_failure;
+
       if (t->collect_md) {
               if (nla_put_flag(skb, IFLA_GRE_COLLECT_METADATA))
                       goto nla_put_failure;

@@ -1084,6 +1092,7 @@ static const struct nla_policy ipgre_policy[IFLA_GRE_MAX + 1] = {
       [IFLA_GRE_ENCAP_SPORT]  = { .type = NLA_U16 },
       [IFLA_GRE_ENCAP_DPORT]  = { .type = NLA_U16 },
       [IFLA_GRE_COLLECT_METADATA]     = { .type = NLA_FLAG },
+       [IFLA_GRE_IGNORE_DF]    = { .type = NLA_U8 },

};

You could just copy the approach used for COLLECT_METADATA and skip
the need to pass any data and instead just use a flag.

I looked at doing that, but decided not to because there’s no way to turn it back off once you turn it on that way.

-Philip

quoted

static struct rtnl_link_ops ipgre_link_ops __read_mostly = {

diff --git a/net/ipv4/ip_tunnel.c b/net/ipv4/ip_tunnel.c
index d8f5e0a..95649eb 100644
--- a/net/ipv4/ip_tunnel.c
+++ b/net/ipv4/ip_tunnel.c

@@ -682,7 +682,7 @@ void ip_tunnel_xmit(struct sk_buff *skb, struct net_device *dev,
       }

       df = tnl_params->frag_off;
-       if (skb->protocol == htons(ETH_P_IP))
+       if (skb->protocol == htons(ETH_P_IP) && !tunnel->ignore_df)
               df |= (inner_iph->frag_off&htons(IP_DF));

       max_headroom = LL_RESERVED_SPACE(rt->dst.dev) + sizeof(struct iphdr)

Looking at this it appears that you should probably make the feature
mutually exclusive from the pmtudisc feature.  It wouldn't make much
sense to set the tunnel to ignore the inner DF bit and then turn
around and set it because you are enforcing path MTU discovery on the
tunnel itself.

From: "Philip Prindeville" <redacted>
Date: Tue,  7 Jun 2016 13:48:46 -0600

From: Philip Prindeville <redacted>

In the presence of firewalls which improperly block ICMP Unreachable
(including Fragmentation Required) messages, Path MTU Discovery is
prevented from working.

A workaround is to handle IPv4 payloads opaquely, ignoring the DF bit--as
is done for other payloads like AppleTalk--and doing transparent
fragmentation and reassembly.

Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Philip Prindeville <redacted>

I think the feedback that this feature should be mutually exclusive
to pmtudisc is valid and you should address that.

On Jun 11, 2016, at 12:02 AM, David Miller [off-list ref] wrote:

From: "Philip Prindeville" <redacted>
Date: Tue,  7 Jun 2016 13:48:46 -0600

quoted

From: Philip Prindeville <redacted>

In the presence of firewalls which improperly block ICMP Unreachable
(including Fragmentation Required) messages, Path MTU Discovery is
prevented from working.

A workaround is to handle IPv4 payloads opaquely, ignoring the DF bit--as
is done for other payloads like AppleTalk--and doing transparent
fragmentation and reassembly.

Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Philip Prindeville <redacted>

I think the feedback that this feature should be mutually exclusive
to pmtudisc is valid and you should address that.

Yes, understood.  Looking into this.

-Philip